# Coding Agent: Best AI coding agent, model and effort per task

> Coding Agent · https://coding-agent.josedacosta.net/en · data of 2026-11-11 · José DA COSTA (https://www.josedacosta.net)
> Version française: https://coding-agent.josedacosta.net/fr.md

For every coding task: which agent, model and reasoning effort to use with your subscription, how to work with it and when to switch — to get the best result before your quota runs out.

## Scope: subscriptions only

This page evaluates coding agents used through a subscription — Claude Pro/Max/Ultra, ChatGPT Plus/Pro for Codex, Google AI Pro/Ultra for Antigravity, SuperGrok, Kimi, GLM Coding Plan, Qwen Token Plan for Qwen Code, MiMo Token Plan for MiMo Code. Pay-per-token usage with an API key is excluded from its reasoning. API prices appear only because they provide a secondary estimate of relative consumption, not a guaranteed conversion to subscription quota.

## What is a coding agent?

An AI that works inside your project: it reads the files, edits the code, runs commands and tests, and iterates until the task is done, with you reviewing. The same agent is often reachable in many ways — a terminal (CLI), a desktop app, an IDE extension, a web service, remotely from a phone or another browser, inside GitHub and CI/CD (pull request reviews, agents assigned to issues, GitHub Actions), from team tools (Slack, Teams, Linear, Jira…) or through an SDK in your own scripts: Codex, for example, runs in Codex CLI, the ChatGPT desktop app, the Codex IDE extension, Codex cloud, Codex Remote, Codex code review, Slack and the Codex SDK. Below, the most used agents, most used first; this page covers them as long as they are used through a subscription.

| Agent | Usage at work | CLI (terminal) | Desktop app | IDE extension | Web / chat | Remote & mobile | GitHub & CI/CD | Team tools | SDK & scripts |
|---|---|---|---|---|---|---|---|---|---|
| Claude Code — Anthropic | 39%, npm 12,482,566/wk | ✓✓ Claude Code CLI (12.5 M npm downloads a week) | ✓ Claude Code Desktop | ✓✓ Claude Code for VS Code (27.5 M VS Code installs, 4.9 M JetBrains downloads) | ✓ Claude Code on the web | ✓ Remote Control, Claude app (iOS, Android) | ✓✓ Claude Code GitHub Actions, GitLab CI/CD (used by 24,081 GitHub repositories) | ✓ Claude Code in Slack | ✓✓ Claude Agent SDK, claude -p (12.1 M npm downloads a week) |
| GitHub Copilot — GitHub | 21%, npm 1,562,659/wk | ✓✓ GitHub Copilot CLI (1.6 M npm downloads a week) | — | ✓✓ GitHub Copilot extension (78.7 M VS Code installs, 55.0 M JetBrains downloads) | ✓ Copilot Chat (github.com/copilot) | ✓ GitHub Mobile | ✓✓ Copilot code review, Copilot cloud agent (60 M+ code reviews, more than 1 in 5 on GitHub) | ✓ Slack, Microsoft Teams, Jira, Linear, Azure Boards | ✓✓ GitHub Copilot SDK, copilot -p (1.5 M npm downloads a week) |
| Codex — OpenAI | 16%, npm 23,140,547/wk | ✓✓ Codex CLI (23.1 M npm downloads a week) | ✓ ChatGPT desktop app | ✓✓ Codex IDE extension (15.4 M VS Code installs) | ✓ Codex cloud | ✓ Codex Remote (ChatGPT mobile app) | ✓✓ Codex code review (GitHub, GitLab preview), Codex GitHub Action (100k+ pull requests reviewed a day (Oct. 2025)) | ✓ Slack, Linear | ✓✓ Codex SDK, codex exec (2.1 M npm downloads a week) |
| Cursor — Cursor | 12% | ✓ Cursor CLI | ✓✓ Cursor (editor) (the Cursor editor: 12% of developers at work) | ✓ JetBrains (ACP) | ✓ Cloud Agents | ✓ Cursor for iOS | ✓✓ Bugbot (2 M+ pull requests reviewed a month) | ✓ Slack, Linear | ✓ Cursor SDK, headless CLI |
| Junie — JetBrains | 9% with JetBrains AI | ✓ Junie CLI | — | ✓✓ Junie in JetBrains IDEs (35.5 M JetBrains downloads) | — | — | ✓ GitHub, GitLab | — | — |
| OpenCode — Anomaly | 7%, npm 2,958,079/wk | ✓✓ OpenCode Terminal (3.0 M npm downloads a week) | ✓ OpenCode Desktop | ✓✓ OpenCode Extensions (1.2 M VS Code installs) | — | ✓ opencode web (network access) | ✓ GitHub, GitLab | — | ✓✓ OpenCode SDK, opencode run (38.4 M npm downloads a week) |
| Antigravity — Google | 6% | ✓ Antigravity CLI | ✓✓ Antigravity IDE, Antigravity 2.0 (the Antigravity editor: 6% of developers at work) | ✓ Antigravity IDE Extensions | — | ✓ Remote Control | — | — | ✓ Antigravity SDK |
| Qwen Code — Alibaba (Qwen) | npm 68,764/wk | ✓ Qwen Code CLI | ✓ Qwen Code Desktop | ✓ Qwen Code Companion | — | ✓ Channels: Telegram, WeChat, QQ | ✓ Qwen Code GitHub Action | ✓ Channels: DingTalk, WeCom, Feishu | ✓ Qwen Code SDK, headless mode |
| Grok Build — xAI | npm 60,608/wk | ✓ Grok Build CLI | — | ✓ via ACP | — | — | — | — | ✓ Headless mode |
| Kimi Code — Moonshot AI | npm 47,815/wk | ✓ Kimi Code CLI | ✓ Kimi Code Desktop | ✓ Kimi Code for VS Code | — | — | — | — | ✓ Kimi Agent SDK |
| ZCode / Claude Code — Z.ai | — | ✓ Claude Code | ✓ ZCode | ✓ Claude Code IDE Plugin | — | ✓ Remote Control, WeChat bot | — | ✓ Feishu bot | — |

✓✓ heavily used on that surface: at least 1 M CLI or SDK downloads a week, 1 M IDE-extension installs, 10,000 repositories using the GitHub action or 1 M pull requests reviewed a month; desktop, web, remote and team tools publish no figures, so only the editor the survey measures (Cursor, Antigravity) gets one there.
✓ available on that surface.

Usage: share of professional developers using the agent at work, May–July 2026 (JetBrains Developer Ecosystem Survey 2026, 15,000+ developers). npm: weekly downloads of the agent's CLI, October 4 – 10, 2026. Downloads include CI runs and automatic updates, so they only order the agents that the survey does not report. Qoder and MiMo Code are left out: fewer than 12,000 weekly downloads each. The gray names are the vendors' own names for each surface; the web column counts hosted services only, not local browser interfaces.

## Current recommendations (2026-11-11)

- **Advanced planning and architecture**: Claude Code · Claude Fable 5.2 · high or Claude Code · Claude Opus 5.5 (default) · xhigh
- **Everyday coding**: Claude Code · Claude Opus 5.5 (default) · medium (default)
- **Debugging**: Claude Code · Claude Opus 5.5 (default) · high
- **Front-end and web UI**: Claude Code · Claude Opus 5.5 (default) · medium (default)
- **Long autonomous agent runs**: Claude Code · Claude Opus 5.5 (default) · xhigh or Codex · GPT-6.1 Sol (default) · xhigh
- **High volume, subagents, quick edits**: Codex · GPT-6.1 Terra · medium (default) or Claude Code · Claude Haiku 5.5 · medium (default)
- **Algorithms, proofs and competitive programming**: Codex · GPT-6 Astra · max or Codex · GPT-6.1 Sol (default) · max
- **Scientific and numerical code**: Claude Code · Claude Fable 5.2 · high or Claude Code · Claude Opus 5.5 (default) · xhigh
- **SVG illustrations and icons as code**: Codex · GPT-6 Astra · xhigh or Claude Code · Claude Opus 5.5 (default) · high
- **Charts, dashboards and diagrams as code**: Claude Code · Claude Opus 5.5 (default) · medium (default)
- **3D, games and creative coding as code**: Claude Code · Claude Opus 5.5 (default) · medium (default) or Codex · GPT-6 Astra · xhigh

## What to use, when

Pick what you are about to do: you get the provider, model, reasoning effort and subscription to use, how long to stay on it and what to switch to next, how many agent-hours your quota gives — with the evidence behind each choice.

### Advanced planning and architecture

Writing a development plan, a spec, a migration strategy, an architecture review.

Examples: Write a development plan before a feature · Design a migration (framework, database, API version) · Architecture or security review of a codebase · Turn a vague brief into a spec and task list

> No benchmark measures the quality of a plan or a spec. The pick rests on general reasoning (AA Intelligence Index) and on LibraryDesignBench, which grades how well a model designs a library's API. Security reviews are a tie on CWE-bench v2 (250 tasks: Opus 5.5 67.1%, Gemini 4 Argon 65.0%, GPT-6 Astra 64.2%, ±5 points; Fable 5.2 and Gemini 4 Carbon not ranked yet).

- **Recommended**: Claude Code · Claude Fable 5.2 · high — Claude Max ($100) or Ultra ($200): Fable may use up to 50% of the weekly limit (on Pro, paid credits only). ≈ 2.9 agent-hours per 5-hour window · ≈ 14 per week on Claude Max 5x (central estimate).
- **Recommended**: Claude Code · Claude Opus 5.5 (default) · xhigh — Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 1.7 agent-hours per 5-hour window · ≈ 13 per week on Claude Pro (central estimate).
- **Stay on it**: Only for the plan: one session, until the plan or `SPEC.md` is written, reviewed and approved.
- **Then**: Stop there. Implement in a fresh session with Opus 5.5 at medium (default) — see Everyday coding — or hand the plan to a long run with `/goal`. Staying at xhigh to implement costs 2.5× more per task (AA index: $6.05 vs $2.45), and implementing with Fable 5.2 would spend its weekly cap on the part a cheaper model does well.

**How to work**

1. Type `/plan <description>` and set `/effort xhigh` (Opus 5.5), or `/model fable` then `/effort high` on Max or Ultra (Claude Code sets Fable 5.2 to medium by default): the agent reads and proposes, nothing is edited until you approve.
2. End the spec with an end-to-end check that proves the feature works, then have a fresh session or subagent review the plan before implementing.
3. For a large feature, ask Claude to interview you and write the result to `SPEC.md`; plans are also saved in `~/.claude/plans`.
4. Edit the plan with `Ctrl+G`, then implement in a fresh session at `/effort medium`.
5. In Codex: `/plan`, with `plan_mode_reasoning_effort` set in `config.toml`, and keep a `PLANS.md` referenced from `AGENTS.md`.

**Why**

- A plan is one long reasoning pass, so effort pays off: Opus 5.5 scores 57.4 on the AA Intelligence Index (v5) at xhigh, above Claude Sonnet 5.5 at max (54.5 for $10.02) and just below its own max (59.1), for 40% less than max ($6.05 vs $10.04 per task). It works on Claude Pro.
- Planning is where Claude Fable 5.2 earns its quota on Max and Ultra: first on the AA index (61.8 at high for $9.08; 64.6 at max), first on FrontierCode 2.0 at xhigh (49.1%; 45.3% at high, level with Opus 5.5 at max, 45.5%, for less than its cost) and on LiveBench 2.0 overall (84.9). A plan is short in tokens but decides everything after it, so a few Fable turns pay off; then hand the work to Opus 5.5, which costs about 1.4 times less quota per hour.
- Opus 5.5 is still first on LibraryDesignBench (48.9 ±0.9), which grades the API a model designs for a library, ahead of Claude Fable 5.1 (47.5) and GPT-6 Astra (45.1) — and of the real production library (46.6); Fable 5.2 is not ranked there yet.
- Plan once at high (Fable 5.2) or xhigh (Opus 5.5), max only for the plan that really matters, then execute the steps at medium (Opus 5.5). Fable 5.2 at max is rarely worth it: +2.7 points on the AA index over high for nearly 3 times the cost ($26.54 vs $9.08).

**Alternatives**

- Codex · GPT-6 Astra · xhigh: Codex: 55.3 on the AA index at $3.66; max adds only +1.5 points for +70% cost.
- Antigravity · Gemini 4 Carbon · high (default): Antigravity on Google AI Ultra: 60.0 on the AA index at high for $7.76 per task (at list price; half until 2026-11-30), above Opus 5.5 at xhigh (57.4), with few hallucinations. On Google AI Pro, Gemini 4 Argon at high: 50.4 on the AA index and first on LiveBench 2.0 reasoning (95.1, against 92–93 for Astra, GPT-6.1 Sol and Opus 5.5, one run); one planning session fits Pro's window.

Sources: Artificial Analysis model leaderboard · Snorkel LibraryDesignBench 1.0.0 · CWE-bench v2 · LiveBench · FrontierCode leaderboard data · Introducing Claude Fable 5.2 and Claude Mythos 5.2 · Claude Fable models on your plan · Gemini 4 Carbon is now available · Antigravity models · Claude Code best practices · Codex best practices

### Everyday coding

Features, simple fixes, reviews, small refactors — the work you do all day. A bug you cannot explain has its own card: Debugging.

Examples: Fix a test broken by your own change · Add an endpoint, a form, a CLI option · Review a pull request · Small refactor in a few files

- **Recommended**: Claude Code · Claude Opus 5.5 (default) · medium (default) — Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).
- **Stay on it**: All day: medium, Claude Code's default, is the setting to leave on. One session per task (`/clear` between tasks), and end a session after two failed corrections.
- **Then**: A task that resists (a subtle change, two failed attempts): `/effort high` for that task only, or `ultrathink` for one turn, then back to medium. A bug you cannot explain: Debugging. A change you cannot describe in one sentence: plan it first (Advanced planning). Many small mechanical edits: High volume.

**How to work**

1. No plan needed when you can describe the change in one sentence; otherwise type `/plan <task>` — more reliable than `Shift+Tab`, which now starts from auto mode.
2. Stay on medium (the default). For one hard turn, add `ultrathink` to the prompt: since 2.1.311 it runs that turn at xhigh in the terminal.
3. Give the agent a check it can run (tests, build, lint) and ask for the command and its output, not a claim. After two failed corrections, `/rewind` or `/clear` and re-prompt with what you learned.
4. Before the pull request: `/code-review` looks for bugs in a fresh subagent; `/simplify` cleans up the diff but does not look for bugs.
5. In Codex: `/model` → GPT-6.1 Sol at medium (its default, low, scores far lower); structure the prompt as Goal, Context, Constraints, Done when.

**Why**

- The best quality for the quota on routine work. Benchmarks are made of hard tasks: there, high adds about 4 points over medium (CursorBench 4.0 52.5% → 56.0%, Terminal-Bench 4.0 (AA) 52.5% → 56.6%, FrontierCode 2.0 30.1% → 34.5%) for about 35% more quota, so about a quarter fewer agent-hours. On everyday tasks that medium already solves, that gain does not show; keep high for the tasks that resist.
- It is the only top model Claude Pro includes without credits, and Claude Code's default model and effort since 2026-09-22. Anthropic's own SWE-bench Pro subset: 92.8% solved at $0.22 per solved task, versus 92.3% at $1.19 for Fable 5.1 at its default.
- Claude Fable 5.2 at medium scores a little higher (CursorBench 54.8%, Terminal-Bench 4.0 (AA) 55.1%, FrontierCode 2.0 37.6%) for more per task ($4.66 vs $2.91 on CursorBench), uses Fable's weekly cap (50% on Max and Ultra) and needs paid credits on Pro: see the alternative.

**Alternatives**

- Codex · GPT-6.1 Sol (default) · medium: Codex on ChatGPT Plus: level with Opus 5.5 at medium on FrontierCode 2.0 (29.6% vs 30.1%) for $0.49 per task against $1.70, and 48.0% on Terminal-Bench 4.0 (AA) against 52.5%. Set medium: its default, low, drops to 30.8% on Terminal-Bench 4.0 (AA). Plus allows about 15–160 GPT-6.1 Sol messages per 5 hours. For UI and visual work prefer Opus 5.5: GPT-6.1 Sol trails it on Arena WebDev (1743 vs 1814) and Design Arena (1354 vs 1381).
- Antigravity · Gemini 4 Carbon · medium: Antigravity on Google AI Ultra (not Pro), or GitHub Copilot Pro+: first on CursorBench 4.0 (60.1% at medium for $8.49) and 40.0% on FrontierCode 2.0, ahead of Opus 5.5 at medium for about three times the cost per task. Set medium: its default, high, adds nothing on FrontierCode 2.0 (37.5%) and low is no cheaper. Its API price doubles on 2026-12-01 ($4 / $20).
- Claude Code · Claude Fable 5.2 · medium (default): On Claude Max or Ultra, for the hardest tasks of the day, not for everything: at medium it scores above Opus 5.5 at high on FrontierCode 2.0 (37.6% vs 34.5%) and on the AA index (56.4 vs 51.8) for $3.10–3.80 per task. It counts against Fable's own cap (50% of the weekly limit). Never `/fast` on a subscription: fast mode is billed as paid credits only ($16 / $80 per million tokens).
- Claude Code or ZCode · GLM-5.3 · max (default): Budget fallback only, when no other subscription fits: GLM Coding Plan Lite ($18/month, 2,000 credits per 5 hours) inside Claude Code; DeepSWE 69.0% at max, but far behind on CursorBench 4.0 (42.6%) and FrontierCode 2.0 (17.8%). At max it burns about 2,900 credits per busy hour, so Lite gives roughly 40 minutes of agent time per 5 hours (twice that off-peak).

Sources: Artificial Analysis model leaderboard · FrontierCode leaderboard data · Arena Code: WebDev leaderboard · Design Arena: Code overall · Optimizing for cost and intelligence · Model configuration · DeepSWE v1.1 leaderboard · Introducing GPT-6 Sol and Luna · Introducing GPT-6.1 Sol · Artificial Analysis Coding Agent Index · Codex pricing · GLM Coding Plan · CursorBench 4.0 · What's new in Gemini 4 Argon (Gemini API) · What's new in Gemini 4 Carbon (Gemini API) · Gemini 4 Carbon is now available · Models and pricing for GitHub Copilot · Introducing Claude Fable 5.2 and Claude Mythos 5.2 · Claude Fable models on your plan · Antigravity models · Antigravity plans · Claude Code best practices · Claude Code commands · Claude Code permission modes

### Debugging

A bug you cannot explain yet: find the cause, reproduce it, fix it and prove the fix. Diagnosis needs more reasoning than writing code.

Examples: A test that fails only in CI, or only sometimes · A stack trace or a crash in production · A regression after an upgrade · A bug that spans several services or files

- **Recommended**: Claude Code · Claude Opus 5.5 (default) · high — Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 1.8 agent-hours per 5-hour window · ≈ 14 per week on Claude Pro (central estimate).
- **Stay on it**: Until the cause is found and the fix is proven by a test, then `/effort medium` again: high costs about 27% more per task than medium, too much to leave on all day.
- **Then**: Still stuck: add `ultrathink` to the next prompt (one turn at xhigh). Stuck after two attempts: `/clear`, then Claude Fable 5.2 at xhigh (Max or Ultra) with what you learned. Once fixed: back to Everyday coding.

**How to work**

1. Start with `/effort high`; add `ultrathink` to a single prompt to run just that turn at xhigh.
2. Give the full error, the stack trace and the exact command that fails, then ask for a reproduction (a failing test) before any fix.
3. Ask for the cause in one sentence and the evidence (logs, a bisect, a value printed) before the agent edits code; reject a fix that only silences the symptom.
4. Done when the reproduction test passes and the rest of the suite still does; then `/code-review` on the diff.
5. In Codex: `/model` → GPT-6.1 Sol at xhigh for the diagnosis, then back to medium.

**Why**

- Diagnosis is a hard task, where effort shows. Terminal-Bench is mostly diagnosis in a terminal (failing builds, broken services, data to recover); on the Artificial Analysis run, Opus 5.5 gains 4 points from medium to high (52.5% → 56.6%) for $4.04 → $5.12 per task, about 27% more quota.
- Why not xhigh for the whole session: 3 more points (59.6%) for $8.78, 72% more again, and nothing more on CursorBench 4.0 (56.0% at high and xhigh). `ultrathink` gives xhigh to the one turn that needs it. Max adds nothing over xhigh (59.6%).

**Alternatives**

- Claude Code · Claude Fable 5.2 · xhigh: On Claude Max or Ultra, for the bug Opus 5.5 could not solve, even with `ultrathink`: the strongest Claude model at diagnosis — Terminal-Bench 4.0 (AA) 61.6%, FrontierCode 2.0 49.1%, AA index 62.2, and 65.2% on the Vals AI Terminal-Bench run (at high) — for $14.83–20.81 per task. Max scores lower (59.6%) for $31.59. It counts against Fable's cap (50% of the weekly limit).
- Antigravity · Gemini 4 Carbon · high (default): Antigravity on Google AI Ultra, or GitHub Copilot Pro+: 65.2% on Terminal-Bench 4.0 (AA) at high for $10.05 per task, ahead of every Claude and OpenAI model on that run, and 60.4% on CursorBench 4.0. Its API price doubles on 2026-12-01.
- Codex · GPT-6.1 Sol (default) · xhigh: Codex on ChatGPT Plus or Pro: 54.0% on Terminal-Bench 4.0 (AA) at xhigh for $1.03 per task, a fifth of Opus 5.5 at high, for a few points less; max adds 2 points (56.1%, and 60.9% on the official Terminal-Bench 4.1 board) for 75% more.

Sources: Artificial Analysis model leaderboard · CursorBench 4.0 · Vals AI benchmarks · Terminal-Bench 4.1 leaderboard · Introducing Claude Fable 5.2 and Claude Mythos 5.2 · Claude Fable models on your plan · Introducing GPT-6.1 Sol · Codex pricing · What's new in Gemini 4 Carbon (Gemini API) · Gemini 4 Carbon is now available · Model configuration · Claude Code best practices · Claude Code commands

### Front-end and web UI

Landing pages, components, dashboards — where humans judge the result.

Examples: Build a landing page or marketing site · Create UI components and design systems · Dashboards and data visualization · Make an existing page responsive and accessible

- **Recommended**: Claude Code · Claude Opus 5.5 (default) · medium (default) — Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).
- **Stay on it**: For pages, components and tweaks: medium, the default, already leads Design Arena, whose Opus 5.5 entries run at a default-like effort.
- **Then**: For the one page that must stand out (a landing page, a showcase): `/effort max` while you build it, the setting behind the Arena WebDev lead, then back to medium. Max burns about 35% more quota per hour (effort factor 1 vs 0.74).

**How to work**

1. Start with `claude --chrome` (subscription login required): the agent opens the page, reads console errors, screenshots its result and compares it with your mockup; `/verify` checks the running app.
2. Name the styles to avoid (for example “no cream background, no pill buttons”) rather than asking to “avoid a generic AI look”.
3. Keep the dev server running and iterate with small prompts.
4. Use plan mode only for multi-page features or a design system.
5. In Antigravity, `/browser` checks the layout in a sandboxed Chrome.

**Why**

- Top of Arena WebDev (1814.3 ±11.7, blind human votes, cutoff 2026-11-10), shared with Claude Fable 5.2 at xhigh (1835.8 ±22.4 on only 994 votes, ranks 1–2), and clearly ahead of Claude Sonnet 5.5 xhigh (1772.5), Gemini 4 Carbon high (1768.2, pre-release), GPT-6 Astra max (1764.4) and GPT-6.1 Sol max (1743.0). First on Arena Image-to-WebDev too (1751.9 ±19.2), level with Claude Sonnet 5.5 (1744.1).
- Design Arena agrees without separating the leaders: level with GPT-6 Astra at xhigh on its overall code ranking (1381 vs 1387), first on the agentic full-stack board (1408, 73% of 318 battles won, ahead of Astra at 1342) and second on the agentic frontend board (1319, behind Astra at max, 1343). Claude Fable 5.2 is only ninth on the overall code ranking (1347) and takes about twice as long per generation.
- Medium for everyday front-end work, max only for a showcase page: Arena WebDev measures Opus 5.5 at max only, while Design Arena states no effort for Claude and its Opus 5.5 entries run about as fast as Claude Sonnet 5.5, a default-like effort — and they already lead the full-stack board.

**Alternatives**

- Claude Code · Claude Fable 5.2 · xhigh: On Claude Max or Ultra, for the one page that must impress: highest Arena WebDev score (1835.8 ±22.4), but on 994 votes, within the margin of Opus 5.5, ninth on Design Arena's overall code ranking (1347) and about twice as slow per generation. It counts against Fable's 50% weekly cap: go back to Opus 5.5 once the page exists.
- Claude Code · Claude Sonnet 5.5 · xhigh: Third on Arena WebDev (1772.5 ±9.9), second on Arena Image-to-WebDev (1744.1) and ahead of Opus 5.5 on Design Arena UI components (1386 vs 1370, on 459 battles), for about 40% of Opus 5.5 max's cost per AA index task ($4.09 vs $10.04).
- Kimi Code · Kimi K3.1 Pro · max: Through the Kimi Code plans (Plus caps its context at 262K tokens): 1687.7 ±9.8 on Arena WebDev, the best open-weight model there, second on Design Arena UI components (1393 over 324 battles, behind GPT-6 Astra at 1399) and third on the agentic full-stack board (1330). Slow: about 15–19 minutes per generation on Design Arena.
- Qwen Code · Qwen3.8 Max · xhigh (default): 1675.8 ±6.2 on Arena WebDev, just below Kimi K3.1 Pro, and 1286 on Design Arena frontend, above Claude Sonnet 5.5 (1281) and GPT-6.1 Sol (1266), in Qwen Code on the Qwen Token Plan from $6.

Sources: Arena Code: WebDev leaderboard · Arena Code: Image-to-WebDev · Design Arena: Code overall · Design Arena: Website · Design Arena: UI components · Design Arena: Mobile apps (agents) · Design Arena: agentic web dev · Design Arena: data visualization · Artificial Analysis model leaderboard · Introducing Claude Fable 5.2 and Claude Mythos 5.2 · Claude Fable models on your plan · Token Plan (Personal Edition) · Kimi Code models · Claude Code in Chrome · Prompting Claude Opus 5.5

### Long autonomous agent runs

Multi-hour refactors, migrations, terminal-heavy work where the agent runs on its own.

Examples: Migrate a whole codebase to a new framework · Multi-hour refactor across hundreds of files · Set up CI, Docker and infrastructure from scratch · Let the agent work alone on a long task list

> The evidence splits by kind of task, so choose by subscription. On SWE-Marathon (long tasks, 160 trials each), Gemini 4 Carbon, Sonnet 5.5, Opus 5.5, GPT-6.1 Sol and GPT-6 Astra are within noise (42–54%). Reverse engineering of binaries favors OpenAI (Vals SRE-bench, a security benchmark: Astra 56.9%, GPT-6.1 Sol 50.8%, Opus 5.5 33.6%), and no benchmark measures infrastructure or DevOps work as such; building a program from scratch favors Claude (Vals ProgramBench: Fable 5.2 19.0%, Opus 5.5 18.5%, Astra 5.5%). On LiveBench 2.0's long-horizon suite of 2026-11-06, Fable 5.2 at xhigh leads (82.9) ahead of Astra (75.6), Opus 5.5 (72.8) and GPT-6.1 Sol (65.6), on one run and at $160.92 per task. The AA Coding Agent Index measures Opus 5.5 and GPT-6.1 Sol at max only.

- **Recommended**: Claude Code · Claude Opus 5.5 (default) · xhigh — Claude Max ($100) or Ultra ($200) for long sessions. One agent never hits the 5-hour limit (≈ 8.6 agent-hours per window) · ≈ 82.2 agent-hours per week on Claude Max 5x (central estimate).
- **Recommended**: Codex · GPT-6.1 Sol (default) · xhigh — ChatGPT Plus ($20) or Pro. ≈ 3.7 agent-hours per 5-hour window · ≈ 26.3 per week on ChatGPT Plus (Codex) (central estimate).
- **Stay on it**: For the whole run, until the `/goal` condition is met — hours if needed. Check the milestones instead of watching every step.
- **Then**: When the goal is met, end the run, review in a fresh context (`/code-review`), then go back to medium (Everyday coding) for follow-ups. If the run fails on the same error twice, stop it and re-plan instead of letting it burn quota.

**How to work**

1. Plan first with `/plan`, then hand over the objective with `/goal <condition>` (Claude Code and Codex). Make the condition checkable (“tests pass”, “every endpoint migrated”): a separate evaluator re-checks it after each turn.
2. Stay in auto mode (the default) so the run never waits on a prompt, and write in `AGENTS.md` / `CLAUDE.md` to stop only when blocked or before anything destructive. Never `--dangerously-skip-permissions` outside a sandbox.
3. Effort: `/effort xhigh` in Claude Code; `model_reasoning_effort = 'xhigh'` for GPT-6.1 Sol in Codex (its default is low).
4. Split independent parts: `/batch <instruction>` or `claude --worktree <name>` in Claude Code; worktrees or cloud tasks in Codex.
5. `ultracode` (or “use a workflow”) orchestrates dozens of agents but uses much more quota: try it on one directory first.
6. Finish with a fresh-context review: `/code-review` (alias `/review`) in Claude Code, `/review` in Codex.

**Why**

- On the AA Coding Agent Index, which runs each model in its vendor's agent, Opus 5.5 at max in Claude Code is first (61.7 for $11.42 per task), ahead of Gemini 4 Carbon (59.4) and Claude Sonnet 5.5 (59.3); GPT-6.1 Sol at max in Codex scores 54.1 for $1.50, lower but for about an eighth of the cost. On the official tbench.ai leaderboard (Terminal-Bench 4.1, the 2026-11-09 regrade of 4.0), Opus 5.5 at max is first (65.76%, $14.30 per trial) and GPT-6.1 Sol at max (60.91%) edges GPT-6 Astra (60.30%) for a fifth of its cost per trial ($1.92 vs $9.90).
- Opus 5.5: max adds nothing over xhigh on AA's Terminal-Bench 4.0 run (both 59.6%) for 49% more cost; Anthropic recommends xhigh for tasks over 30 minutes.
- GPT-6.1 Sol burns about half of GPT-6 Sol's quota per busy hour in the AA Coding Agent Index, so ChatGPT Plus is enough to start. xhigh: Terminal-Bench 4.0 (AA) 54.0% for $1.03 per task, the score of GPT-6 Astra at high for a quarter of the cost. Max adds about 2 points (56.1%) for 75% more, which a multi-hour run on Plus feels: keep max for a short, hard run.

**Alternatives**

- Claude Code · Claude Sonnet 5.5 · max: Third on the AA Coding Agent Index in Claude Code, 59.3 at $13.30 per task against 61.7 at $11.42 for Opus 5.5 max, but ahead on Terminal-Bench 4.0 (AA), 63.6% against 59.6%, and second on the official tbench.ai leaderboard (63.94%). Since its cache reads were halved on 2026-10-07 it should burn about 40% less quota per busy hour than Opus 5.5 (derived).
- Codex · GPT-6 Astra · high: For the hardest long runs in Codex on ChatGPT Pro: first on Vals SRE-bench (reverse engineering of binaries) at max (56.9%, GPT-6.1 Sol 50.8%, Opus 5.5 33.6%), second on FrontierSWE (65.5, behind Fable 5.2) and on LiveBench 2.0's long-horizon suite at max (75.6, one run). Elsewhere GPT-6.1 Sol matches it for far less (AA Coding Agent Index at max: 54.7 for $7.04, against 54.1 for $1.50). On the official tbench.ai leaderboard high already reaches 60.0%.
- Antigravity · Gemini 4 Carbon · high (default): On Google AI Ultra in Antigravity (`/goal` included): second single model on the AA Coding Agent Index (59.4 for $7.74 per task at list price, about $15.60 per busy agent-hour, half that until 2026-11-30), first on SWE-Marathon (53.75%, within noise of the others) and the best worst run on FrontierSWE (56.2; mean 61.2 for $61.93 per trial, the cheapest of the leaders). Weaker at shell steps in Arena's agent sessions; give `/swarm` an agent limit. Google AI Pro has no Carbon, and Gemini 4 Argon there gives only about 2 agent-hours per 5-hour window.
- Claude Code · Claude Fable 5.2 · high: Only for the hardest build of the week, on Claude Max or Ultra: first on FrontierSWE (69.9, against 62.3 for Opus 5.5) and on Vals ProgramBench (19.0% at high), and on LiveBench 2.0 long-horizon at xhigh (82.9). But it costs about 1.7 times Opus 5.5 per trial on FrontierSWE ($168.78 vs $98.87), Fable may use at most 50% of the weekly limit, and its system card reports work left unfinished in 6–13% of sessions: give `/goal` a checkable condition.

Sources: Artificial Analysis model leaderboard · Terminal-Bench 4.1 leaderboard · Artificial Analysis Coding Agent Index · Introducing Claude Sonnet 5.5 · Introducing Claude Haiku 5.5 · Introducing GPT-6.1 Sol · Effort · Claude Code /goal · Claude Code workflows · Claude Code permission modes · SWE-Marathon v1.1 · Vals AI: SRE-bench · Vals AI: ProgramBench · FrontierSWE V2 · LiveBench · Antigravity models · Gemini 4 Carbon and /swarm · Introducing Claude Fable 5.2 and Claude Mythos 5.2 · Claude Fable 5.2 and Claude Mythos 5.2 System Card · Claude Fable models on your plan

### High volume, subagents, quick edits

Many small, well-defined tasks: the best quality a small budget buys, without spending the quota of a frontier model on them.

Examples: Rename, reformat, add types across many files · Write docstrings, tests or translations in bulk · Subagents that search or summarize code · Quick one-line fixes

- **Recommended**: Codex · GPT-6.1 Terra · medium (default) — ChatGPT Plus ($20). One agent never hits the 5-hour limit (≈ 6.6 agent-hours per window) · ≈ 47.4 agent-hours per week on ChatGPT Plus (Codex) (central estimate).
- **Recommended**: Claude Code · Claude Haiku 5.5 · medium (default) — Claude Pro ($20), Max ($100) or Ultra ($200). One agent never hits the 5-hour limit (≈ 16.1 agent-hours per window) · ≈ 122.3 agent-hours per week on Claude Pro (central estimate).
- **Stay on it**: As long as the tasks stay mechanical and fully specified: renames, types, docstrings, tests from a clear pattern. Test the prompt on 2–3 files, then fan out.
- **Then**: As soon as a task needs judgment (a design choice, an unclear bug), go back to Everyday coding. Only when the Terra quota runs out on trivial edits, GPT-6 Luna at max costs a little less, but scores far lower (Terminal-Bench 4.0 (AA) 12.6% against 32.3%). Never Claude Fable 5.2 for bulk work or subagents: it would empty its 50% weekly cap on edits a small model does as well.

**How to work**

1. A cheaper model rather than a frontier model at low effort: GPT-6.1 Terra in Codex (`/model`), or Claude Haiku 5.5 in Claude Code (`/model haiku`), both at their default, medium.
2. Fan out: `/batch <instruction>` or a `claude -p` loop in Claude Code; ask Codex to spawn subagents.
3. Test the prompt on 2–3 files before fanning out, and restrict tools with `--allowedTools`.
4. Every subagent draws on the same subscription quota: give them a lower effort (`effort: low`) or a smaller model (`model: haiku`, Claude Haiku 5.5) in `.claude/agents/<name>.md`; in Codex, `agents.default_subagent_reasoning_effort`.
5. Put the shared rules in `AGENTS.md` / `CLAUDE.md` so every run follows them.

**Why**

- Quality first, then price, for easy tasks: bulk edits are mechanical and fully specified, so the default effort is enough. At medium, GPT-6.1 Terra scores 32.3% on Terminal-Bench 4.0 (AA) for $0.28 per task, 38.2 on the AA index for $0.17 and 42.0 on the AA Coding Agent Index for $0.39: ahead of Claude Haiku 5.5 at medium (15.2%, 24.6) and far ahead of GPT-6 Luna at max (12.6%, 27.9, 31.2).
- When a batch item is harder than expected, raise the effort for that item only: Terra at high reaches 41.9% on Terminal-Bench 4.0 (AA) for $0.44, and Haiku 5.5 at xhigh 44.3% on CursorBench 4.0 for $0.56 (36.9% at medium).
- In Claude Code, Haiku 5.5 is the small model to use: it holds up on long tasks (SWE-Marathon at max: 50.0%), and Anthropic pitches it for subagents, summaries and compaction; `model: haiku` in `.claude/agents/<name>.md` makes it the subagent model.

**Alternatives**

- Claude Code · Claude Opus 5.5 (default) · low: Inside Claude Code, Anthropic recommends low for subagents: 87.4% on its SWE-bench Pro subset at $0.12 per solved task.
- MiMo Code · MiMo-V2.6-Pro · high (default): Cheapest capable option: 34.8% on Terminal-Bench 4.0 (AA) at $0.44 per task and 29.9 on the AA index at $0.26 — ahead of GLM-5.3-Flash (32.8% at $0.89; 27.8 at $0.58) — and 57.8% on the Vals Index for $0.43, in MiMo Code on the MiMo Token Plan from $6. AA does not state the effort of these MiMo runs (assumed high).
- Codex · GPT-6 Luna · max: Last resort in Codex when the Terra quota runs out: ChatGPT Plus allows about 350–3,000 Luna messages per 5 hours against 35–285 for Terra, but Luna at max trails on every independent board (Terminal-Bench 4.0 (AA) 12.6%, AA index 27.9, SWE-Marathon 28.1%). Keep it for trivial, fully specified edits; never below max (high: 4.5% on Terminal-Bench 4.0 (AA)).

Sources: Introducing GPT-6 Sol and Luna · Introducing GPT-6.1 Sol · Introducing GPT-6.1 Terra · Codex pricing · Codex models · Optimizing for cost and intelligence · DeepSWE v1.1 leaderboard · GLM Coding Plan overview · Claude Code best practices · Claude Code costs · Codex best practices · MiMo Token Plan · Artificial Analysis model leaderboard · Introducing Claude Haiku 5.5 · Claude Haiku 5.5 (Claude Platform docs) · Model configuration · CursorBench 4.0 · Artificial Analysis Coding Agent Index · SWE-Marathon v1.1 · FrontierCode leaderboard data

### Algorithms, proofs and competitive programming

A very difficult algorithm, an optimization problem, a proof, a competitive-programming-style task.

Examples: Optimize a hard algorithm (graphs, dynamic programming) · Competitive-programming-style problems · Prove or check a mathematical property · A heuristic for a hard optimization (scheduling, routing)

> OpenAI models lead every algorithm board, but Opus 5.5 is measured on ALE-Bench at high only, and Claude Fable 5.2 and Gemini 4 Carbon are not yet on ALE-Bench, Vals IOI or WeirdML v3.

- **Recommended**: Codex · GPT-6 Astra · max — ChatGPT Pro ($100 / $200 / $500). ≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate).
- **Recommended**: Codex · GPT-6.1 Sol (default) · max — ChatGPT Plus ($20) or Pro. ≈ 3.7 agent-hours per 5-hour window · ≈ 26.3 per week on ChatGPT Plus (Codex) (central estimate).
- **Stay on it**: Only for the hard core — the algorithm, the proof, the optimization: one focused session.
- **Then**: Once the core is correct and tested, go back to Everyday coding for the surrounding code, I/O and tests.

**How to work**

1. Ask first for a brute-force reference and tests, then for the optimized algorithm, and let the agent compare both.
2. In Codex: `/model` → GPT-6 Astra, `model_reasoning_effort = 'max'`; on ChatGPT Plus, GPT-6.1 Sol at max.
3. For a proof, ask for each step to be checked by a test or a small program (or a proof assistant) rather than by the model's own reading.
4. On a Claude plan: Opus 5.5 at `/effort xhigh`, or `ultrathink` for a single hard turn.

**Why**

- Quality first: GPT-6 Astra at max leads ALE-Bench (2951, against 2147 for Opus 5.5 at high), scores 100% on Vals IOI (Opus 5.5 95.1% ±4.9) and leads WeirdML v3 (42.2 at xhigh, Opus 5.5 31.2).
- GPT-6.1 Sol at max is close for far less, on ChatGPT Plus: second on ALE-Bench (2744 for $0.60, against $2.86 for Astra), 96.9% on Vals IOI for $1.26 per task and second on WeirdML v3 (38.3). GPT-6.1 Terra at max is third on ALE-Bench (2609 for $0.41).

**Alternatives**

- Antigravity · Gemini 4 Argon · high (default): For competitive-programming-style problems in Antigravity on paid Google AI Pro or Ultra: 100% on Vals IOI (level with GPT-6 Astra) for $5.31 per task and first on LiveBench 2.0 reasoning (95.1). Weaker on optimization and machine-learning tasks: 2100 on ALE-Bench at high, 20.3 on WeirdML v3.
- Claude Code · Claude Opus 5.5 (default) · xhigh: If you are on a Claude plan only: 95.1% on Vals IOI and 31.2 on WeirdML v3 at xhigh, behind the OpenAI models; ALE-Bench measures it at high only (2147). Avoid max: Anthropic warns that it can overthink.

Sources: Vals AI benchmarks · ALE-Bench (Epoch AI Benchmarking Hub) · WeirdML v3 (Epoch AI Benchmarking Hub) · LiveBench · Claude Opus 5.5 · Codex pricing

### Scientific and numerical code

Simulations, numerical kernels, data science and research code where a wrong formula silently gives wrong results.

Examples: A physics or finance simulation · Port a numerical kernel (NumPy, Julia, CUDA) · Reproduce a method from a paper · A data-analysis pipeline with statistics

> Two boards, partly disjoint: Terminal-Bench Science (Artificial Analysis) measures every current model, SciCode does not yet measure Claude Fable 5.2.

- **Recommended**: Claude Code · Claude Fable 5.2 · high — Claude Max ($100) or Ultra ($200): Fable may use up to 50% of the weekly limit (on Pro, paid credits only). ≈ 2.9 agent-hours per 5-hour window · ≈ 14 per week on Claude Max 5x (central estimate).
- **Recommended**: Claude Code · Claude Opus 5.5 (default) · xhigh — Claude Max ($100) or Ultra ($200) for long sessions. One agent never hits the 5-hour limit (≈ 8.6 agent-hours per window) · ≈ 82.2 agent-hours per week on Claude Max 5x (central estimate).
- **Stay on it**: For the numerical core: the model, the formulas, the kernel, until it matches a reference result.
- **Then**: Once the core matches its reference, go back to Everyday coding (Opus 5.5 at medium) for I/O, plots and packaging: Fable 5.2 would spend its weekly cap on code a cheaper model writes as well.

**How to work**

1. `/model fable` then `/effort high` on Max or Ultra (Claude Code sets Fable 5.2 to medium by default); on Pro, Opus 5.5 at `/effort xhigh`.
2. Give a reference: a known analytical solution, a small case computed by hand, or the paper's published numbers, and have the agent compare against it after each change.
3. Ask for units, tolerances and invariants (conservation, symmetry) as tests, not as comments.

**Why**

- Quality first: Claude Fable 5.2 leads Terminal-Bench Science (AA) at high (70.0%; 73.3% at max), well ahead of Gemini 4 Carbon (64.8%), GPT-6 Astra (63.3% at max) and Opus 5.5 at xhigh (61.9%). High rather than xhigh (67.6%) or max, which costs far more for 3 points.
- Opus 5.5 is first on SciCode (66.9% at max, 65.0% at xhigh), ahead of Gemini 4 Carbon (64.2%) and about ten points above GPT-6 Astra (56.5%), and works on Claude Pro without Fable's weekly cap. xhigh rather than max: Terminal-Bench Science (AA) 61.9% at xhigh against 59.0% at max.

**Alternatives**

- Antigravity · Gemini 4 Carbon · high (default): Antigravity on Google AI Ultra: second on Terminal-Bench Science (AA) (64.8%) and close to Opus 5.5 on SciCode (64.2%). On Google AI Pro, Gemini 4 Argon is weaker here (SciCode 61.8%, Terminal-Bench Science 52.9%).
- Codex · GPT-6.1 Sol (default) · max: Codex on ChatGPT Plus: 58.1% on Terminal-Bench Science (AA) for far less per task, a few points behind the Claude and Google leaders.

Sources: Artificial Analysis evaluation details · Artificial Analysis: Terminal-Bench-Science · Terminal-Bench-Science 0.1 (Snorkel mirror of the official board) · Artificial Analysis model leaderboard · Claude Opus 5.5 · Introducing Claude Fable 5.2 and Claude Mythos 5.2 · Claude Fable models on your plan · Antigravity models

### SVG illustrations and icons as code

Icons, logos and illustrations written as SVG code. Code generation only: this is not image generation (no raster pictures from a prompt). Charts and diagrams have their own card.

Examples: A consistent icon set for an app, in SVG · An illustration or a wordmark as a single SVG file · A logo with a few variants (dark, light, square) · A small animated SVG for a landing page

> Design Arena is the only large blind-vote board: it labels GPT-6 Astra as xhigh but gives no effort for Claude models (on that board a higher effort adds about 20 Elo, far less than Astra's lead), and it publishes a standard error again. Playcode's benchmark draws one sample per model and effort, judged by eye. Claude Fable 5.2, Gemini 4 Carbon, GPT-6.1 Sol and GPT-6.1 Terra are not ranked on Design Arena SVG yet (about 200 battles per category are needed): Fable 5.2 may do better, but nothing measures it yet.

- **Recommended**: Codex · GPT-6 Astra · xhigh — ChatGPT Plus ($20, 5–45 Astra messages per 5 hours) or Pro ($100 / $200 / $500). ≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate).
- **Recommended**: Claude Code · Claude Opus 5.5 (default) · high — Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 1.8 agent-hours per 5-hour window · ≈ 14 per week on Claude Pro (central estimate).
- **Stay on it**: While you create the assets — an icon set, an illustration, a logo — iterating with renders until it looks right.
- **Then**: For small edits to an existing SVG (colors, sizes), your everyday agent is enough.

**How to work**

1. Ask for a file, not inline code: one `.svg` per asset with a `viewBox`, no embedded raster images, and a fixed grid (for example 24 × 24 for icons) so a set stays consistent.
2. Close the loop visually: have the agent render the file in a browser, take a screenshot and fix what it sees before you review.

**Why**

- First general-purpose model in the Design Arena SVG arena: Elo 1421 ±9 over 1,808 blind battles (68.8% won), about 90 points above Gemini 4 Argon (1329), Kimi K3.1 Pro (1326) and Claude Opus 5.5 (1321 ±10, 1,261 battles): about seven standard errors, the only large-sample signal for SVG illustrations. Only the SVG-specialized Arrow 2 Telos, in no subscription, scores higher.
- Playcode's MacBook SVG benchmark (34 models, one drawing per model and effort, judged by eye, updated 2026-11-09) now ranks Claude Opus 5.5 first, Claude Fable 5.2 second and GPT-6 Astra third. Opus 5.5's best drawing is at max on the Batch API, which a subscription cannot use: streamed, max returns an empty file; high draws a correct laptop for $0.40 and xhigh a more detailed one for $1.72, four times the cost — keep xhigh for a showcase illustration. Fable 5.2 draws its best laptop at medium ($3.91) and nothing at xhigh or max. Every Astra effort draws a correct laptop ($0.98 at xhigh).

**Alternatives**

- Codex · GPT-6 Sol · max: The cheapest route on ChatGPT Plus (15–150 messages per 5 hours, against 5–45 for Astra): sixth on Playcode's MacBook SVG benchmark, with a correct product shot for $0.39 at max, about a sixth of Astra's $2.29 at max (GPT-6.1 Sol, fifth, draws as much for $0.21 at xhigh). Thirteenth on Design Arena SVG (1302 over 893 battles, effort not stated), close to Claude Opus 5.5.
- Antigravity · Gemini 4 Argon · high (default): On Google AI Pro ($19.99) in Antigravity: the best general-purpose model on Design Arena SVG after GPT-6 Astra (1329 over 628 battles, 61.1% won), just above Opus 5.5. Effort not stated by the board; Antigravity defaults to High.

Sources: Design Arena: SVG · MacBook SVG Benchmark: 34 AI Models, Ranked by the Laptop They Draw · Managing usage with GPT-6 Astra in Work and Codex · Codex pricing · Prompting Claude Opus 5.5 · Antigravity models · Antigravity plans

### Charts, dashboards and diagrams as code

Data visualizations, dashboards and architecture or sequence diagrams written as code (ECharts, D3, Mermaid, PlantUML). Code generation only, not image generation.

Examples: An interactive chart or data dashboard · An architecture or sequence diagram in Mermaid · A flowchart of a process, kept next to the code · Turn a CSV into a readable report with charts

> Charts have one large blind-vote board, Design Arena data visualization; no benchmark compares models on diagram languages (Mermaid, PlantUML, draw.io, Excalidraw), so for diagrams the pick rests on the chart board and on cost. Claude Fable 5.2 and Gemini 4 Carbon are not ranked on it yet.

- **Recommended**: Claude Code · Claude Opus 5.5 (default) · medium (default) — Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).
- **Stay on it**: While you build the chart or the first version of the diagram, iterating with renders until it reads right.
- **Then**: For later edits (a label, a new node, a color), your everyday agent is enough: a Mermaid diagram is ordinary text once it exists.

**How to work**

1. Close the loop visually: have the agent render the chart in a browser (or `npx @mermaid-js/mermaid-cli` for Mermaid), take a screenshot and fix what it sees before you review.
2. Keep diagrams as text (Mermaid in Markdown) next to the code they describe, so the agent can update them with each change.
3. Give the real data (or a sample) and say what the reader must see at a glance; ask for an accessible table next to each chart.

**Why**

- Claude Opus 5.5 leads Design Arena data visualization (1390 ±14 over 700 blind battles), ahead of Claude Sonnet 5.5 (1362, on only 160 battles), Kimi K3 (1348), Claude Fable 5.1 (1342), Gemini 4 Argon (1333) and GPT-6 Astra (1325, 10th): about three standard errors ahead of Astra, the leader for SVG illustrations.
- Medium, the default: Design Arena states no effort for Claude, and charts and diagrams are ordinary code once the data is known. Raise to high for a complex dashboard; avoid max, which on Playcode's SVG benchmark returns an empty file after 128,000 tokens of thinking.

**Alternatives**

- Claude Code · Claude Sonnet 5.5 · xhigh: Second on Design Arena data visualization (1362), on only 160 battles so far: a provisional rank, for less quota per hour than Opus 5.5.
- Kimi Code · Kimi K3 · high (default): Third in the data-visualization arena (1348), on a much cheaper Kimi Code plan. On the same plans, Kimi K3.1 Pro is only sixteenth there (1306).
- Antigravity · Gemini 4 Argon · high (default): On Google AI Pro ($19.99) in Antigravity: seventh on data visualization (1333). Effort not stated by the board; Antigravity defaults to High.

Sources: Design Arena: data visualization · MacBook SVG Benchmark: 34 AI Models, Ranked by the Laptop They Draw · Prompting Claude Opus 5.5 · Antigravity models · Antigravity plans

### 3D, games and creative coding as code

Three.js scenes, WebGL, browser games, generative art and animation, written as code. Code generation only: this is not image or 3D-model generation.

Examples: A complex interactive 3D scene in Three.js · A playable browser game with physics · Shaders, particles and procedural animation · Data art with p5.js or D3

> No current board measures Three.js or WebGL scenes (the ZeroEval Three.js arena stopped in August 2026 with no current model). The closest is Design Arena's game-dev board, which publishes no standard error: GPT-6 Astra at max leads (1466, 1,244 battles) ahead of Astra at xhigh (1457) and Claude Opus 5.5 (1429) — a lead, not a decisive one — then GPT-6.1 Sol (1409), Kimi K3.1 Pro (1406), Claude Fable 5.2 (1404, on only 396 battles) and Claude Sonnet 5.5 (1394) follow close together. Choose by subscription. Design Arena states no effort for Claude models.

- **Recommended**: Claude Code · Claude Opus 5.5 (default) · medium (default) — Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).
- **Recommended**: Codex · GPT-6 Astra · xhigh — ChatGPT Pro ($100 / $200 / $500). ≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate).
- **Stay on it**: While you build the scene and its interactions, at the default effort, medium.
- **Then**: For tuning values, colors and small fixes, `/effort medium` (default). Textures and icons as SVG: see SVG illustrations and icons.

**How to work**

1. Plan the scene structure first (plan mode), then build it step by step with the dev server running.
2. Check every iteration visually: with `claude --chrome` the agent runs the scene, reads console errors and screenshots the result to compare with what you expect.
3. Build at medium (the default); raise to `/effort high` for shaders and physics that resist, or `ultrathink` for one hard turn. Avoid max: no board shows a gain for 3D, and it can spend the whole 128,000-token output limit on one large file.

**Why**

- Games: the best Claude model on Design Arena game dev (1429, 2,666 battles), behind GPT-6 Astra at max (1466) and xhigh (1457), and ahead of Claude Fable 5.2 (1404), on browser games written as code.
- Visual web work: top of Arena WebDev (1814.3 ±11.7 at max, shared with Claude Fable 5.2; Astra at max 1764.4 ±8.6) and level with Astra at xhigh on Design Arena's overall code ranking (1381 vs 1387). Neither measures 3D as such.
- Why medium: the Opus 5.5 entry on game dev states no effort, and it takes about as long per generation as Claude Sonnet 5.5 (483 s vs 486 s), less than Astra at xhigh (603 s) and far less than Astra at max (1,133 s), which points to the default effort, medium. The only max entry, Arena WebDev, is web-app work.

**Alternatives**

- Codex · GPT-6 Astra · max: The choice in Codex: first on Design Arena game dev at max (1466 over 1,244 battles; 1457 at xhigh) and fifth on Arena WebDev at max (1764.4). Max is the most quota-hungry setting in Codex: on ChatGPT Plus (5–45 Astra messages per 5 hours) keep it for the hard parts.
- Claude Code · Claude Sonnet 5.5 · xhigh: 1394 on Design Arena game dev (2,008 battles), 35 points below Opus 5.5, and third on Arena WebDev at xhigh (1772.5, behind Opus 5.5 and Fable 5.2), for about 40% less quota per busy hour at the same effort (derived).

Sources: Design Arena: Game dev · Arena Code: WebDev leaderboard · Design Arena: Code overall · Managing usage with GPT-6 Astra in Work and Codex · MacBook SVG Benchmark: 34 AI Models, Ranked by the Laptop They Draw · Claude Code in Chrome

# Practical coding-agent workflows

> https://coding-agent.josedacosta.net/en#workflows · Content checked on 2026-09-29

Choose a task for its method, commands for your agent and a request you can adapt. Each command names its surface; availability depends on your version and enabled skills.

## Your first task

Use this for a small, visible change you can check yourself.

Before you start: A working starting point, one desired result, and a saved copy or Git commit.

1. Open the relevant project and confirm the agent can see the right files.
2. Describe one change and its acceptance criteria. Ask for a short plan if the scope is unclear.
3. Inspect the diff, run the relevant checks, then try the result yourself.

### Useful commands

- **Claude Code CLI · Codex CLI** — `/diff`: Inspect changes before accepting them.
- **Claude Code CLI** — `/run`: Launch and operate the app to observe the result.

- [Claude Code commands](https://code.claude.com/docs/en/commands) — Anthropic; Verified 2026-09-29
- [Codex developer commands](https://learn.chatgpt.com/docs/developer-commands) — OpenAI; Verified 2026-09-29
- [Claude Code: run and verify your app](https://code.claude.com/docs/en/skills) — Anthropic; Verified 2026-09-29

### Request to adapt

```text
In [project], implement [one change]. Preserve [constraints]. Success means [observable result]. First inspect the relevant files. Run [checks] and report the result, changed files and remaining limitations.
```

Done when: For a form: submit valid and invalid values, reload and verify persistence if required, then try a narrow screen.

If it fails: If the scope grows, save the current state and split the next change into a separate task.

- [Codex best practices](https://learn.chatgpt.com/guides/best-practices) — OpenAI; Verified 2026-10-10
- [Claude Code best practices](https://code.claude.com/docs/en/best-practices) — Anthropic; Verified 2026-09-28

## Prepare a project

Use this when the agent repeatedly misses project conventions.

Before you start: The real build, test and run commands, project boundaries and a few recurring constraints.

1. Run the documented commands once so the instructions describe reality.
2. Put durable conventions in the instruction file supported by your agent. Keep task details in the request.
3. Ask the agent to identify the instructions it loaded and the checks relevant to this change.

### Useful commands

- **Codex CLI** — `/init`: Create an AGENTS.md scaffold, then adapt its rules to the repository.
- **Claude Code CLI** — `/init`: Set up CLAUDE.md; preserve existing instructions and imports.
- **Claude Code CLI** — `/run-skill-generator`: Record the project launch procedure in a reusable skill.

- [Codex developer commands](https://learn.chatgpt.com/docs/developer-commands) — OpenAI; Verified 2026-09-29
- [Claude Code commands](https://code.claude.com/docs/en/commands) — Anthropic; Verified 2026-09-29
- [Claude Code: run and verify your app](https://code.claude.com/docs/en/skills) — Anthropic; Verified 2026-09-29

### Request to adapt

```text
Read this project’s agent instructions. Report the run/test commands and the files relevant to [task]. Propose a short addition only for a recurring rule that is missing. Verify the command before documenting it.
```

Done when: A new session can find the commands and constraints without repeating the previous conversation.

If it fails: Remove stale or conflicting rules. This repository uses AGENTS.md and a CLAUDE.md import; that convention is not universal.

- [Codex best practices](https://learn.chatgpt.com/guides/best-practices) — OpenAI; Verified 2026-10-10
- [Claude Code best practices](https://code.claude.com/docs/en/best-practices) — Anthropic; Verified 2026-09-28

## Write a verifiable request

Use this when “make it work” leaves too much interpretation.

Before you start: The user journey, current behaviour, expected behaviour and constraints.

1. Name the starting state and one representative example.
2. State the success conditions, including an error path when it matters.
3. Specify how to verify them and what evidence the agent should report.

### Useful commands

- **Claude Code CLI · Codex CLI** — `/plan`: Clarify scope and checks before coding when the task remains ambiguous.

- [Claude Code commands](https://code.claude.com/docs/en/commands) — Anthropic; Verified 2026-09-29
- [Codex developer commands](https://learn.chatgpt.com/docs/developer-commands) — OpenAI; Verified 2026-09-29

### Request to adapt

```text
Starting from [state], when [action], I expect [result]. Today [actual result]. Respect [constraints]. Check [normal case] and [error case]. Report observed results separately from assumptions.
```

Done when: Another person can follow the checks and decide whether the request is satisfied.

If it fails: If the expected result is still ambiguous, decide it before asking for more implementation.

- [Codex prompting](https://learn.chatgpt.com/docs/prompting) — OpenAI; Verified 2026-09-29

## Understand an existing project

Use this before changing an unfamiliar feature or reviewing a correction.

Before you start: A concrete question, entry point or user action.

1. Trace one journey from input to output.
2. Ask for file and symbol references and distinguish observations from hypotheses.
3. Open the cited files or run a small example to check the explanation.

### Useful commands

- **Claude Code CLI · Codex CLI** — `/plan`: Request an analysis of the flow in plan mode before implementation.
- **Codex CLI** — `/mention`: Attach the specific file to investigate to the request.

- [Claude Code commands](https://code.claude.com/docs/en/commands) — Anthropic; Verified 2026-09-29
- [Codex developer commands](https://learn.chatgpt.com/docs/developer-commands) — OpenAI; Verified 2026-09-29

### Request to adapt

```text
Explain how [user action] reaches [result]. Cite the relevant files and symbols, describe the state changes and identify what you could not verify. Do not modify files yet.
```

Done when: You can locate the code responsible for the behaviour and reproduce one example.

If it fails: Narrow the question to one route or function if the explanation becomes a repository-wide inventory.

- [Codex best practices](https://learn.chatgpt.com/guides/best-practices) — OpenAI; Verified 2026-10-10

## Reproduce and fix a bug

Use this for a reproducible failure, including a regression.

Before you start: Expected versus actual behaviour, reproduction steps, useful error text and environment.

1. Reproduce the failure before changing code.
2. Ask for a cause supported by evidence and a focused correction.
3. Repeat the reproduction and relevant regression checks after the correction.

### Useful commands

- **Claude Code CLI · Codex CLI** — `/diff`: Inspect the fix and spot changes unrelated to the bug.
- **Claude Code CLI** — `/verify`: Exercise the app after the fix to check its behaviour.

- [Claude Code commands](https://code.claude.com/docs/en/commands) — Anthropic; Verified 2026-09-29
- [Codex developer commands](https://learn.chatgpt.com/docs/developer-commands) — OpenAI; Verified 2026-09-29
- [Claude Code: run and verify your app](https://code.claude.com/docs/en/skills) — Anthropic; Verified 2026-09-29

### Request to adapt

```text
Reproduce [steps]. Expected: [result]. Actual: [error]. Find evidence for the cause, apply a focused fix, then repeat the reproduction and [relevant checks]. Report any check you could not run.
```

Done when: The original failure no longer occurs and relevant checks pass. A passing unrelated test is not sufficient.

If it fails: If reproduction fails, preserve logs and environment details. Do not pile speculative fixes onto the same session.

- [Codex prompting](https://learn.chatgpt.com/docs/prompting) — OpenAI; Verified 2026-09-29
- [Claude Code best practices](https://code.claude.com/docs/en/best-practices) — Anthropic; Verified 2026-09-28

## Verify before delivery

Use this before publishing, merging or handing work to someone else.

Before you start: The intended behaviour, diff, relevant checks and representative data.

1. Inspect the diff, then run a code review using your agent’s command below.
2. Run checks that exercise the changed behaviour; inspect the real UI for visual changes.
3. Record actual outcomes, untested cases and any remaining action.

### Useful commands

- **Claude Code CLI** — `/code-review`: Look for bugs in the current diff. /review is an alias from version 2.1.223.
- **Codex CLI · Desktop** — `/review`: Choose local changes or a base branch; review reports findings without editing files.
- **Codex — terminal** — `codex review --uncommitted`: Run a non-interactive review of uncommitted changes, including untracked files.
- **Claude Code CLI** — `/verify`: Build and run the app, then observe the result. May save the procedure as a project skill.
- **Claude Code CLI** — `/simplify`: Optional: applies code cleanup. Inspect the diff and rerun checks afterwards.

- [Claude Code review](https://code.claude.com/docs/en/code-review) — Anthropic; Verified 2026-09-29
- [Codex code review](https://learn.chatgpt.com/docs/code-review) — OpenAI; Verified 2026-10-10
- [Codex developer commands](https://learn.chatgpt.com/docs/developer-commands) — OpenAI; Verified 2026-09-29
- [Claude Code: run and verify your app](https://code.claude.com/docs/en/skills) — Anthropic; Verified 2026-09-29

### Request to adapt

```text
Review this change against [acceptance criteria]. Check [main journey], [error path] and [constraints]. Run the relevant tests. Report findings with file references, checks performed and remaining uncertainty.
```

Done when: The checks have observable results and someone can distinguish verified behaviour from assumptions.

If it fails: Fix material findings and repeat the affected checks. A second agent’s approval is not a substitute for evidence.

- [Codex best practices](https://learn.chatgpt.com/guides/best-practices) — OpenAI; Verified 2026-10-10
- [Claude Code best practices](https://code.claude.com/docs/en/best-practices) — Anthropic; Verified 2026-09-28

## Resume or hand off a task

Use this before a context reset, quota interruption or change of agent.

Before you start: Goal, constraints, changed files, attempts, observed results and next useful step.

1. Save work and identify the exact branch or working copy.
2. Write a short handoff with decisions, failed attempts and commands already run.
3. Ask the new session to verify the saved state before continuing.

### Useful commands

- **Claude Code CLI · Codex CLI** — `/compact`: Summarize the conversation to continue the same task with less context.
- **Claude Code CLI · Codex CLI** — `/resume`: Reopen a saved session in the same agent. To switch agents, use the handoff below.

- [Claude Code commands](https://code.claude.com/docs/en/commands) — Anthropic; Verified 2026-09-29
- [Codex developer commands](https://learn.chatgpt.com/docs/developer-commands) — OpenAI; Verified 2026-09-29

### Request to adapt

```text
Prepare a handoff: objective, constraints, current branch/files, decisions, checks and results, failed attempts, outstanding questions and the next concrete step. Exclude secrets and unnecessary conversation history.
```

Done when: A new session can resume from a verified state without repeating failed attempts.

If it fails: If state differs, reconcile files and checks first. Do not assume the handoff proves the current state.

- [Claude Code best practices](https://code.claude.com/docs/en/best-practices) — Anthropic; Verified 2026-09-28
- [Claude Code costs](https://code.claude.com/docs/en/costs) — Anthropic; Verified 2026-09-28

## Distribute work between agents

Use this for independent tasks with clear boundaries and a person responsible for integration.

Before you start: Separate deliverables, owned files or isolated worktrees, and shared acceptance criteria.

1. Separate independent investigation or implementation tasks. Avoid concurrent edits to the same files.
2. Give each agent context, scope and checks; designate an integrator.
3. Review the combined result and run integration checks after bringing changes together.

### Useful commands

- **Claude Code CLI** — `/batch`: For large repetitive changes: describe the work, approve the plan, then distribute edits across worktrees. Budget for quota and integration review.
- **Codex CLI** — `/fork`: Create another conversation to explore an approach. This alone does not create an isolated worktree.

- [Claude Code commands](https://code.claude.com/docs/en/commands) — Anthropic; Verified 2026-09-29
- [Codex developer commands](https://learn.chatgpt.com/docs/developer-commands) — OpenAI; Verified 2026-09-29

### Request to adapt

```text
Own [scope/files] for [deliverable]. Coordinate before changing shared files. Validate with [checks]. Return changes, results and integration notes. The integrator will verify the combined behaviour.
```

Done when: Each contribution is verified and the combined change works. Parallel agents consume additional quota.

If it fails: Serialize work if tasks depend on each other or conflict resolution costs more than parallelism saves.

- [Claude Code worktrees](https://code.claude.com/docs/en/worktrees) — Anthropic; Verified 2026-09-28
- [Claude Code costs](https://code.claude.com/docs/en/costs) — Anthropic; Verified 2026-09-28

## Instructions, skills, hooks and connections

Use this when a recurring task needs a reusable method or a specific external tool.

Before you start: A demonstrated recurring need and the agent’s official documentation for that surface.

1. Use project instructions for durable rules, a request for one task, and a skill for a repeatable procedure.
2. Use a hook for a supported lifecycle action, MCP for a tool connection, and a subagent for a bounded delegated task. Support differs between agents.
3. Check permissions and data sent to external services. Test on a small case and document how to disable or update the mechanism.

### Useful commands

- **Claude Code CLI · Codex CLI** — `/mcp`: Inspect connected servers and tools before using them.
- **Codex CLI** — `/skills`: Select an installed skill for the next request.
- **Claude Code CLI** — `/hooks`: Inspect the project’s hook configuration.

- [Claude Code commands](https://code.claude.com/docs/en/commands) — Anthropic; Verified 2026-09-29
- [Codex developer commands](https://learn.chatgpt.com/docs/developer-commands) — OpenAI; Verified 2026-09-29

### Request to adapt

```text
For recurring [need], compare the simplest supported mechanism: instruction, skill, hook or tool connection. State the supported surface, required permissions, data accessed, verification and maintenance. Install nothing until the choice is clear.
```

Done when: The mechanism solves a recurring need and its permissions, inputs and maintenance are understood.

If it fails: Remove unused configuration. An approval prompt is not a sandbox, and a local interface does not imply local model inference.

- [Codex best practices](https://learn.chatgpt.com/guides/best-practices) — OpenAI; Verified 2026-10-10
- [Agent Security](https://cursor.com/docs/agent/security) — Cursor; Verified 2026-09-29


## What not to do right now

The habits that waste quota or quality with today’s models and plans — each with what to do instead and the figures behind it. Re-checked at every data update.

### Code everything with Claude Fable 5.2, especially at max

- **Instead:** Claude Opus 5.5 at medium (default) for the day's work; Fable 5.2 at high only for planning, scientific code and the hardest build, on Max or Ultra.
- Fable 5.2 leads the general indexes, but max is its worst deal: FrontierCode 2.0 45.7% at max for $55.87 per task, below its own xhigh (49.1% for $17.73); Terminal-Bench 4.0 (AA) 59.6% at max for $31.59, the score of Opus 5.5 at xhigh for $8.78; CursorBench 4.0 60.0% at max for $32.79, level with Gemini 4 Carbon at high (60.4% for $9.22). Fable may use at most 50% of the weekly limit on Max and Ultra and needs paid credits on Pro, so coding everything with it ends the week early. Fable 5.1 does no better (CursorBench 51.8% at max for $17.28, against 56.0% for $3.97 for Opus 5.5 at high).

### Turn on `/fast` with Claude Fable 5.2 on a subscription

- **Instead:** Normal speed within your plan; fast mode only when you accept paying for it.
- Fable 5.2's fast mode (up to 2.5 times faster output) is billed at $16 / $80 per million tokens, twice the normal price, and only as paid usage credits on Claude plans: it never draws on the included quota.

### Leave max on for everyday work

- **Instead:** The agent's default effort for everyday work; high for debugging, xhigh for planning and multi-hour runs; max only where you have measured a gain.
- Opus 5.5: on Terminal-Bench 4.0 (AA) max equals xhigh (59.6%) for 49% more. Claude Fable 5.2: xhigh beats max on FrontierCode 2.0 (49.1% vs 45.7%) for a third of the cost. GPT-6 Astra: xhigh matches or beats max on Terminal-Bench 4.0 and DeepSWE (59.6% vs 59.1%; 74.1% vs 73.2%) for about 30% less. Anthropic warns that max can overthink.

### Drop Gemini 4 Carbon to low to save quota

- **Instead:** Medium for interactive work, high for long runs; for cheap work, a smaller model (Gemini 3.8 Flash).
- On Carbon, low thinks less but takes more steps: AA index 54.2 at low for $7.19 per task against 58.3 at medium for $6.59; CursorBench 4.0 58.6% at low for $11.89 against 60.1% at medium for $8.49. Lower score, same or higher cost.

### Run `ultracode`, a workflow or an unbounded `/swarm` for ordinary tasks

- **Instead:** One session for normal work; keep workflows for a large migration or audit, tried on one directory first.
- A workflow orchestrates dozens of agents, and every agent draws on the same subscription limits: Claude Code warns beyond 25 agents or 1.5M tokens, and Google warns that Antigravity's multi-agent `/swarm` uses quota much faster. Combined with Fable or max, a single run can burn a week of quota.

### Pick Mistral Large 4 for agentic coding on the strength of its launch charts

- **Instead:** Among open-weight models, Kimi K3.1 Pro or GLM-5.3; for cheap volume, GPT-6.1 Terra or Claude Haiku 5.5. Keep Mistral Large 4 for sovereign, self-hosted setups.
- Mistral's 2026-10-29 launch cited 41% on Terminal-Bench 4.0 and 48 on the AA Coding Agent Index; the independent runs now show 24.7% (AA, high), 26.3% (Vals) and 35.4 on the AA Coding Agent Index (30th of 35), against 67.2% and 63.6% on Terminal-Bench 4.0 for Kimi K3.1 Pro. On the Vals Index it scores 56.5, below every other open or low-cost model measured.

### Drop Claude Opus 5.5 to low for real agentic work

- **Instead:** medium (default); keep low for subagents and bulk edits.
- Terminal-Bench 4.0 (AA): low 31.3% vs medium 52.5% — 21 points lost to save $1.97 per task.

### Use GPT-6 Sol, even at max, for long terminal-heavy runs

- **Instead:** GPT-6.1 Sol at xhigh, or GPT-6 Astra at high: ten points higher, for a quarter of the cost or the same cost.
- Terminal-Bench 4.0 (AA): GPT-6 Sol max 43.9% for $4.00 per task, Astra high 54.0% for $4.05, GPT-6.1 Sol xhigh 54.0% for $1.03. On the official tbench.ai leaderboard GPT-6 Sol at max also trails (50.3%, against 60.9% for GPT-6.1 Sol and 60.3% for GPT-6 Astra at max) for more than twice GPT-6.1 Sol's cost per trial ($4.40 vs $1.92), and on Vals SRE-bench (reverse engineering of binaries) it scores 30.5% against 50.8% and 56.9%. GPT-6 Sol stays a good everyday model.

### Draw SVG with Claude Opus 5.5 at max

- **Instead:** Opus 5.5 at high or xhigh — or GPT-6 Astra at xhigh, the first general-purpose model on Design Arena SVG.
- In Playcode's SVG benchmark, Opus 5.5 at max spent the whole 128,000-token output limit thinking and returned no drawing (given 300,000 tokens on the Batch API it finished, which a subscription cannot do), while high drew a correct laptop for $0.40 and xhigh a detailed one for $1.72. Anthropic advises keeping xhigh and max for measured gains.

### Push Grok 4.7 above high (default)

- **Instead:** Stay on high (default).
- xhigh adds 1 point on Terminal-Bench 4.0 (24.7% → 25.8%) for 63% more cost, and 0.3 point on the AA index (35.2 → 35.5) for 26% more. CursorBench gains 2.4 points (43.9% → 46.3%) for 28% more.

### Give Kimi K3 long terminal-heavy tasks

- **Instead:** On the same Kimi Code plans, Kimi K3.1 Pro (`/model`) for terminal work; GLM-5.3 at max (41.9%) on a budget. Keep Kimi K3 for data visualization, where it ranks third.
- Terminal-Bench 4.0 (AA): Kimi K3 scores 12.6% (AA now lists it as deprecated). Kimi K3.1 Pro is first on that board (67.2% at max) and fourth on Vals' run (63.6% at max), but scores only 38.7 on LiveBench 2.0's long-horizon tasks of 2026-11-06 (GPT-6.1 Sol 65.6; one run) and 27.1% on Vals SRE-bench: check it on your own long runs before relying on it.

### Run with `--dangerously-skip-permissions` to go faster

- **Instead:** Auto mode (the default): a classifier screens each action; or a sandbox or container.
- It removes every safeguard on file deletion, shell commands and network access, while auto mode already avoids most prompts.

### Force plan mode on every small change

- **Instead:** Plan only ambiguous or multi-file work with `/plan`.
- Claude Code's and Grok Build's docs both say to skip the plan when the change is obvious or fits in one sentence: for small changes it mostly costs a round trip.

## Habits that pay off in every agent

- **Verification first**: Give the agent a pass/fail check (tests, build, screenshot) and ask for evidence, not claims. Every vendor puts this tip first.
- **Plan only when it helps**: Plan ambiguous, multi-file or unfamiliar work; skip it for one-sentence changes (Claude Code and Grok Build docs). `/plan` is there when you need it.
- **One task per session**: `/clear` between unrelated tasks, `/rewind` instead of piling up corrections, `/btw` for a side question, `/compact <what to keep>` mid-task.
- **Short instruction files**: Keep `AGENTS.md` / `CLAUDE.md` short: add a rule when a mistake repeats, remove what the model already does, and re-test after each new model (`/doctor`).
- **Raise the effort, don't beg**: Start at the agent's default effort; raise it (or `ultrathink` for one turn) instead of writing “think harder”, and keep xhigh and max for tasks where they measurably help.
- **Fresh-context review**: Have code checked by an agent that did not write it (`/code-review`, `/review`), and ask it to report only correctness problems.
- **Parallel work costs quota**: Worktrees, parallel sessions, subagents and workflows multiply the work done, but all draw on the same subscription limits: check the simulator before scaling out.
- **Auto mode, not skip-permissions**: Use auto mode (a classifier screens each action) or a sandbox; never `--dangerously-skip-permissions` in a normal session.

## How long will your subscriptions last?

Enter the subscriptions you have and how you work: you get the agent hours per 5-hour window and per week, and when the quota runs out.

This range is not a statistical confidence interval. Context, task complexity, tools and provider policies affect actual usage. No extra credits or paid overage are assumed. The amounts and burn rates come from the linked subscription sources; personal counters are your assumptions.

An agent-hour is one agent actively working for one hour, not an hour of human availability. Overlapping blocks are concurrent; earlier rows receive quota first. The calendar is an illustrative Monday–Sunday week, without timezone/DST conversion.

Percentages apply to every independent allowance in the row. Separate different counters into separate rows. Unknown weekly/monthly resets do not refill during this week; unspecified five-hour windows start at first use. Model sublimits initially use the same remaining fraction as the overall reserve.

Record separate sessions with the same plan, model, effort and surface. At least three valid samples are required; if the maximum/minimum consumption ratio exceeds three, the general estimate remains. Compare with later sessions before claiming better accuracy.

| Subscription | Monthly price | Agent |
|---|---|---|
| Claude Pro | $20 | Claude Code |
| Claude Max 5x | $100 | Claude Code |
| Claude Ultra | $200 | Claude Code |
| ChatGPT Plus (Codex) | $20 | Codex |
| ChatGPT Pro 100 (Codex) | $100 | Codex |
| ChatGPT Pro 200 (Codex) | $200 | Codex |
| ChatGPT Pro 500 (Codex) | $500 | Codex |
| Google AI Pro (Antigravity) | $19.99 | Antigravity |
| Google AI Ultra 5x (Antigravity) | $99.99 | Antigravity |
| Google AI Ultra 20x (Antigravity) | $199.99 | Antigravity |
| X Subscription Premium (Grok Build) | $30 | Grok Build |
| GLM Coding Plan Lite | $18 | Claude Code / ZCode |
| GLM Coding Plan Pro | $80 | Claude Code / ZCode |
| GLM Coding Plan Max | $168 | Claude Code / ZCode |
| Qwen Token Plan Lite | $6 | Qwen Code |
| Qwen Token Plan Essential | $10 | Qwen Code |
| Qwen Token Plan Standard | $18 | Qwen Code |
| Qwen Token Plan Pro | $68 | Qwen Code |
| MiMo Token Plan Lite | $6 | MiMo Code |
| MiMo Token Plan Standard | $16 | MiMo Code |
| MiMo Token Plan Pro | $50 | MiMo Code |
| MiMo Token Plan Max | $100 | MiMo Code |

## Performance against quota burn

Each line is one model across its reasoning-effort levels. Further left = the task burns less of your quota; higher = better result. The dashed red step line is the efficient frontier: no other model × effort does better for less.

### Terminal-Bench 4.0 (Artificial Analysis run) ⭐⭐⭐

Independent, same harness for every model at every effort level, more than 100 models; the widest controlled run of the reference agentic terminal benchmark.

Agentic terminal tasks run by Artificial Analysis with the same harness for every model, at every effort level.

- Kimi K3.1 Pro · max: 67.1717, $5.2656 per task
- Gemini 4 Carbon · high: 65.1515, $10.053 per task
- Claude Sonnet 5.5 · max: 63.6364, $12.6464 per task
- Gemini 4 Carbon · low: 62.6263, $8.5663 per task
- Claude Fable 5.2 · xhigh: 61.6162, $20.8121 per task
- Gemini 4 Carbon · medium: 60.6061, $8.8204 per task
- Claude Fable 5.2 · high: 59.596, $13.4688 per task
- Claude Fable 5.2 · max: 59.596, $31.5898 per task
- Claude Opus 5.5 · xhigh: 59.596, $8.7784 per task
- Claude Opus 5.5 · max: 59.596, $13.1093 per task
- GPT-6 Astra · xhigh: 59.596, $5.8555 per task
- GPT-6 Astra · max: 59.0909, $8.4981 per task
- Claude Sonnet 5.5 · xhigh: 57.0707, $6.0784 per task
- Gemini 4 Argon · high: 57.0707, $13.5603 per task
- Muse Wonder Full · max: 57.0707
- Claude Opus 5.5 · high: 56.5657, $5.1189 per task
- GPT-6.1 Sol · max: 56.0606, $1.8228 per task
- GPT-6.1 Terra · max: 55.5556, $1.4853 per task
- Claude Fable 5.1 · xhigh: 55.0505, $15.528 per task
- Claude Fable 5.2 · medium: 55.0505, $5.251 per task
- Claude Fable 5.2 · low: 54.5455, $4.3399 per task
- GPT-6.1 Sol · xhigh: 54.0404, $1.0259 per task
- GPT-6 Astra · high: 54.0404, $4.047 per task
- Claude Opus 5.5 · medium: 52.5253, $4.043 per task
- Claude Fable 5.1 · high: 52.0202, $11.4875 per task
- Claude Fable 5.1 · max: 52.0202, $18.773 per task
- Muse Wonder · max: 52.0202, $12.2132 per task
- GPT-6.1 Sol · high: 51.5152, $0.8282 per task
- GPT-6 Astra · medium: 49.4949, $4.4301 per task
- Claude Opus 5 · max: 48.9899, $19.2891 per task
- GPT-6.1 Sol · medium: 47.9798, $0.6138 per task
- GPT-6.1 Terra · xhigh: 46.4646, $0.657 per task
- Claude Fable 5.1 · medium: 44.9495, $8.9758 per task
- Claude Sonnet 5.5 · high: 43.9394, $2.1862 per task
- GPT-6 Sol · max: 43.9394, $4.0034 per task
- GLM-5.3 · max: 41.9192, $8.0642 per task
- GPT-6.1 Terra · high: 41.9192, $0.4375 per task
- GPT-6 Astra · low: 41.9192, $2.2486 per task
- Muse Ember · max: 40.9091, $0.2658 per task
- Claude Fable 5.1 · low: 40.404, $7.1978 per task
- Qwen3.8 Max · xhigh: 38.8889, $18.6761 per task
- Muse Wonder · high: 38.3838, $8.1001 per task
- Tencent Hy4 · high: 37.3737, $1.637 per task
- GPT-5.6 Terra · max: 35.3535, $6.3874 per task
- Muse Ember · xhigh: 35.3535, $0.222 per task
- GLM-5.3 · low: 34.8485, $6.4049 per task
- MiMo-V2.6-Pro · high: 34.8485, $0.4381 per task
- DeepSeek V4.1 Flash 1017 · max: 33.8384, $0.8801 per task
- Ling 3.1 Flash · high: 33.3333
- Muse Spark 1.3 · max: 33.3333, $5.5835 per task
- Step 5 Preview · high: 33.3333, $5.1754 per task
- Claude Haiku 5.5 · max: 32.8283, $0.6519 per task
- GLM-5.3-Flash · max: 32.8283, $0.8892 per task
- Ling 3.1 Flash 1019 · high: 32.8283, $6.5914 per task
- GPT-6.1 Terra · medium: 32.3232, $0.2757 per task
- Claude Opus 5.5 · low: 31.3131, $2.0777 per task
- GPT-6.1 Sol · low: 30.8081, $0.3829 per task
- Claude Sonnet 5.5 · medium: 29.798, $1.5782 per task
- Nanbeige4.3-33B · high: 29.798
- Claude Haiku 5.5 · xhigh: 29.2929, $0.4603 per task
- DeepSeek V4.1 Flash 1017 · high: 28.7879, $0.783 per task
- GPT-6.1 Terra · none: 26.7677, $0.3788 per task
- Grok 4.7 · xhigh: 25.7576, $12.4261 per task
- MiniMax-M3.1-Flash · max: 25.7576, $0.9241 per task
- Qwen3.8-Flash-Next · high: 25.2525
- Grok 4.7 · high: 24.7475, $7.6258 per task
- Mistral Large 4 · high: 24.7475, $4.5925 per task
- Mistral Large 4 · medium: 23.2323, $2.8864 per task
- MiMo-V2.6-Flash · high: 22.7273, $0.2302 per task
- Claude Haiku 5.5 · high: 21.7172, $0.3315 per task
- Claude Sonnet 5.5 · low: 20.7071, $1.2848 per task
- Gemini 3.8 Flash · medium: 19.697, $14.1722 per task
- Gemini 3.8 Flash · high: 19.697, $18.1498 per task
- Grok 4.6 · xhigh: 17.1717, $6.5955 per task
- Grok 4.7 · low: 16.1616, $5.0333 per task
- Claude Haiku 5.5 · medium: 15.1515, $0.1938 per task
- GPT-6.1 Terra · low: 15.1515, $0.1066 per task
- DeepSeek V4 Pro 0813 · max: 14.1414, $3.7552 per task
- Inkling 2 Small · high: 13.1313, $2.4767 per task
- Claude Haiku 5.5 · low: 12.6263, $0.1186 per task
- GPT-5.5 Instant (June 2026) · high: 12.6263
- GPT-6 Luna · max: 12.6263, $0.2362 per task
- Kimi K3 · low: 12.6263, $9.4713 per task
- DeepSeek V4 Flash Vision · max: 12.1212
- GPT-5.6 Luna · max: 11.6162, $0.8493 per task
- Qwen3.8 2.4T A95B · high: 11.1111
- Agnes 2.5 Pro Theta · high: 10.101, $0.7947 per task
- Gemini 3.8 Flash · low: 10.101
- DeepSeek V4.1 Flash 1017 · none: 9.0909, $0.4815 per task
- GPT-6 Luna · xhigh: 8.0808, $0.1982 per task
- Muse Spark Open · xhigh: 8.0808, $4.657 per task
- Qwen3.8 27B · xhigh: 5.5556, $4.5669 per task
- Qwen3.8 27B · medium: 5.0505, $6.5334 per task
- GPT-6 Luna · high: 4.5455, $0.1204 per task
- Gemini 3.1 Pro (preview) · high: 4.0404
- JT-4.1 Flash 236B A21B · high: 3.0303
- Beam 1.0 · high: 3.0303, $0.6254 per task
- GPT-6 Luna · medium: 2.5253, $0.0716 per task
- Qwen3.8 27B · low: 2.5253, $6.9042 per task
- MiniMax-M3 · high: 2.0202, $3.0399 per task
- DeepSeek V4 Pro 0813 · none: 1.5152, $2.5807 per task
- GPT-6 Luna · none: 1.5152, $0.0446 per task
- K2 Horizon 375B A23B · high: 1.5152
- Gemini 3.5 Flash-Lite · high: 1.0101, $1.0951 per task
- Inkling · xhigh: 1.0101, $2.5821 per task
- Inkling Small · high: 1.0101, $0.3508 per task
- K2 Horizon 7B · high: 1.0101
- Kimi K2.7 Code · high: 1.0101
- Quasar 438B · max: 1.0101
- Qwen3.7 Plus · high: 1.0101
- Solar Mini 4 · high: 1.0101
- Command A+ · high: 0.5051
- Hy3 · high: 0.5051
- Muse Glimmer · high: 0.5051, $0.2979 per task
- Nemotron 3.5 Lightning · high: 0.5051, $0.3447 per task
- North Mini Code · high: 0.5051
- Nemotron 3 Ultra 550B A55B · high: 0.5051, $1.7107 per task
- Qwen3.5 9B · high: 0.5051
- Ring-2.6-1T · xhigh: 0.5051
- Trinity Large Thinking · high: 0.5051
- A.X-K2 · high: 0
- Apodex 1.1 · high: 0, $2.0476 per task
- Celeris-1 · none: 0
- Claude Haiku 4.5 · high: 0, $0.9633 per task
- Gemma 4 31B · high: 0
- Gemma 4 E4B · high: 0
- GPT-6 Luna · low: 0, $0.0054 per task
- GPT-OSS 120B · high: 0, $0.0644 per task
- GPT-OSS 20B · high: 0
- Granite 4.2 30B · high: 0
- Granite 4.2 3B · high: 0
- Granite 4.2 8B · high: 0, $0.0399 per task
- HyperNova 60B 2605 · high: 0
- K-EXAONE 2.0 · high: 0
- K2 Horizon 0.9B · high: 0
- K2 Horizon 3.7B · high: 0
- K2 Horizon MoVA 36B A4B · high: 0
- Ling-3.0-flash-Fin · high: 0
- Ling-3.0-flash-VL · high: 0
- Ling 3.0 Tiny · high: 0
- LongCat 2.0 · high: 0
- Mercury 2.5 · high: 0
- MiniCPM5-2B · high: 0
- Ministral 3 14B · none: 0
- Ministral 3 3B · none: 0
- Ministral 3 8B · none: 0
- Mistral Medium 3.5 · high: 0, $1.7048 per task
- Mistral Small 4 · high: 0, $0.0624 per task
- Nemotron 3.5 Bolt 74B-A9B · high: 0
- Nemotron 3.5 Super VL · high: 0, $0.7327 per task
- NVIDIA Nemotron 3 Nano 30B A3B · high: 0
- NVIDIA Nemotron 3 Nano 4B · high: 0
- Qwen3.5 122B A10B · high: 0
- Qwen3.5 397B A17B · high: 0
- Qwen3.6 35B A3B · high: 0
- Qwen3.8 27B · none: 0, $9.3917 per task
- Qwen3 Coder Next · none: 0
- Solar Pro 4 · high: 0

### FrontierCode 2.0 ⭐⭐⭐

Recent, independent, tasks written by open-source maintainers and graded on whether they would merge the change, at every published effort.

Cognition's FrontierCode 2.0 (2026-10-27): 132 coding tasks written by open-source maintainers, graded on mergeability with stricter correctness, design and thoroughness criteria than 1.1; a solution that fails a blocking criterion scores 0. Every model at each published effort, with the cost per rollout at list API prices.

- Claude Fable 5.2 · xhigh: 49.11, $17.734 per task
- Claude Fable 5.2 · max: 45.65, $55.868 per task
- Claude Opus 5.5 · max: 45.45, $12.821 per task
- Claude Fable 5.2 · high: 45.29, $8.402 per task
- Claude Opus 5.5 · xhigh: 43.38, $5.184 per task
- GPT-6 Astra · max: 42.77, $6.676 per task
- GPT-6 Astra · xhigh: 41.41, $4.539 per task
- Claude Sonnet 5.5 · max: 40.96, $19.941 per task
- Gemini 4 Carbon · medium: 40, $4.044 per task
- GPT-6.1 Sol · max: 38.95, $1.029 per task
- GPT-6 Astra · high: 38.31, $4.277 per task
- Claude Fable 5.2 · medium: 37.59, $3.103 per task
- Gemini 4 Carbon · high: 37.48, $4.434 per task
- GPT-6 Astra · medium: 36.75, $3.493 per task
- GPT-6.1 Sol · high: 34.82, $0.647 per task
- Claude Sonnet 5.5 · xhigh: 34.75, $3.315 per task
- Claude Opus 5.5 · high: 34.53, $2.423 per task
- GPT-6 Sol · max: 34.25, $2.3547 per task
- Claude Fable 5.2 · low: 34.2, $2.623 per task
- Claude Fable 5.1 · xhigh: 32.44, $8.68 per task
- GPT-6.1 Sol · xhigh: 32.33, $0.695 per task
- Gemini 4 Carbon · low: 31.88, $4.548 per task
- Claude Fable 5.1 · max: 31.67, $11.457 per task
- GPT-6 Astra · low: 31.11, $2.437 per task
- GPT-6 Sol · xhigh: 30.32, $1.4432 per task
- Claude Opus 5.5 · medium: 30.1, $1.699 per task
- GPT-6.1 Sol · medium: 29.58, $0.492 per task
- Grok 4.7 · xhigh: 28.62, $14.36 per task
- Claude Haiku 5.5 · max: 27.72, $0.366 per task
- GPT-5.6 Sol · max: 27.64, $5.503 per task
- Claude Opus 5 · xhigh: 26.74, $11.093 per task
- Claude Opus 5 · high: 26.72, $8.361 per task
- Claude Opus 5 · max: 26.58, $12.997 per task
- Grok 4.7 · high: 26.55, $11.094 per task
- Gemini 4 Argon · high: 26.3, $6.514 per task
- Grok 4.7 · medium: 25.94, $8.028 per task
- Claude Fable 5.1 · high: 25.67, $5.386 per task
- GPT-6.1 Sol · low: 25.6, $0.341 per task
- GPT-5.6 Sol · high: 24.29, $3.166 per task
- Muse Wonder · high: 24.28, $3.143 per task
- Muse Wonder · xhigh: 24.11, $3.793 per task
- GPT-5.6 Sol · xhigh: 24.08, $3.83 per task
- SWE-2 Flash · max: 24, $0.2024 per task
- Claude Opus 5 · medium: 23.97, $5.217 per task
- SWE-2 · max: 23.78, $1.261 per task
- Gemini 4 Argon · medium: 23.45, $5.688 per task
- SWE-2 Flash · low: 23.17, $0.203 per task
- Muse Wonder · medium: 23.14, $2.534 per task
- SWE-2 · high: 23.06, $0.958 per task
- GPT-6 Sol · high: 22.81, $1.0714 per task
- GPT-5.5 · xhigh: 22.62, $5.2771 per task
- Inkling 2 Small · max: 22.57, $0.584 per task
- Claude Haiku 5.5 · xhigh: 22.28, $0.226 per task
- Inkling 2 Small · high: 22.27, $0.562 per task
- Claude Sonnet 5.5 · high: 21.9, $1.066 per task
- GPT-6 Luna · max: 21.87, $0.174 per task
- GPT-6 Sol · medium: 21.63, $0.8225 per task
- Claude Fable 5.1 · medium: 21.47, $3.891 per task
- Gemini 4 Argon · low: 21.47, $3.977 per task
- Muse Wonder · low: 21.21, $1.843 per task
- Grok 4.7 · low: 20.8, $3.873 per task
- Inkling 2 Small · xhigh: 20.68, $0.615 per task
- Claude Fable 5.1 · low: 20.59, $3.007 per task
- Mistral Large 4 · high: 19.65, $1.549 per task
- Kimi K3.1 Pro · max: 19.64, $3.033 per task
- GPT-5.6 Sol · medium: 19.52, $2.36 per task
- Inkling 2 Small · medium: 19.33, $0.417 per task
- DeepSeek V4.1 Flash 1017 · max: 19.3, $0.384 per task
- GPT-5.5 · high: 19.09, $3.8021 per task
- Claude Opus 5 · low: 18.69, $3.276 per task
- Claude Haiku 5.5 · high: 18.64, $0.139 per task
- GPT-6 Luna · xhigh: 18.44, $0.132 per task
- DeepSeek V4.1 Flash 1017 · high: 18.21, $0.274 per task
- GLM-5.3 · max: 17.81, $7.516 per task
- Kimi K3 · high: 17.74, $4.475 per task
- Inkling 2 Small · none: 17.43, $0.439 per task
- Kimi K3 · max: 17.31, $5.456 per task
- Kimi K3.1 Pro · high: 17.29, $2.358 per task
- GLM-5.3 · high: 16.84, $5.181 per task
- GPT-6 Luna · high: 16.5, $0.115 per task
- SWE-1.7 · high: 16.15, $3.5209 per task
- GPT-6 Luna · medium: 15.97, $0.092 per task
- DeepSeek V4.1 Flash 1017 · low: 15.96, $0.206 per task
- Inkling 2 Small · low: 15.57, $0.286 per task
- SWE-2 · medium: 15.56, $0.442 per task
- Claude Opus 4.8 · max: 14.72, $12.129 per task
- Claude Opus 5.5 · low: 14.72, $0.746 per task
- GLM-5.3 · low: 14.37, $3.189 per task
- Claude Opus 4.8 · xhigh: 14.28, $9.133 per task
- GPT-5.5 · medium: 14.22, $2.9515 per task
- GLM-5.3-Flash · max: 14.18, $0.792 per task
- Claude Sonnet 5.5 · medium: 14.17, $0.482 per task
- Claude Haiku 5.5 · medium: 13.42, $0.097 per task
- Kimi K3.1 Pro · low: 12.86, $1.877 per task
- GPT-5.6 Sol · low: 12.79, $1.451 per task
- GLM-5.3-Flash · high: 12.41, $0.584 per task
- Kimi K3 · low: 11.1, $2.977 per task
- Claude Opus 4.8 · medium: 11.02, $5.588 per task
- Claude Opus 4.8 · high: 10.86, $6.316 per task
- GPT-6 Sol · low: 10.49, $0.4164 per task
- Claude Haiku 5.5 · low: 10.46, $0.065 per task
- Claude Opus 4.8 · low: 9.55, $4.195 per task
- Mistral Large 4 · low: 9.5, $0.669 per task
- Claude Sonnet 5.5 · low: 9.41, $0.369 per task
- Mistral Large 4 · medium: 8.78, $1.031 per task
- GLM-5.3-Flash · low: 7.81, $0.43 per task
- GPT-5.5 · low: 7.2, $1.9753 per task
- Muse Spark Open · xhigh: 6.24, $9.157 per task
- GPT-6 Luna · low: 4.05, $0.023 per task
- Muse Spark Open · high: 3.95, $6.78 per task
- Inkling Small · max: 3.21, $3.684 per task
- Beam 1.0 · max: 2.9, $1.299 per task
- Beam 1.0 · xhigh: 2.43, $1.126 per task
- Beam 1.0 · high: 1.74, $0.612 per task
- Muse Spark Open · medium: 1.63, $1.865 per task
- SWE-1.6 · high: 1.34, $1.3592 per task
- Inkling Small · high: 1.27, $1.366 per task
- Beam 1.0 · medium: 0.83, $0.315 per task
- Muse Spark Open · low: 0.42, $0.593 per task
- Beam 1.0 · low: 0, $0.142 per task

### Terminal-Bench 4.1 (official leaderboard) ⭐⭐⭐

The official leaderboard of the most cited agentic coding benchmark: 330 trials per entry with confidence intervals, each model in its vendor's own agent.

Official tbench.ai leaderboard, Terminal-Bench 4.1 (2026-11-09): the 66 tasks of 4.0 with hardened verifiers, the 4.0 submissions regraded from their saved runs. 330 trials per agent × model × effort, with confidence intervals. Each entry uses its vendor's agent (Claude Code, Codex, Grok Build…).

- Claude Opus 5.5 · max: 65.76, $14.2989 per task
- Claude Sonnet 5.5 · max: 63.94, $22.2232 per task
- GPT-6.1 Sol · max: 60.91, $1.9219 per task
- GPT-6 Astra · max: 60.3, $9.9005 per task
- GPT-6 Astra · high: 60, $6.877 per task
- GPT-6 Astra · xhigh: 59.7, $7.1228 per task
- Claude Fable 5.1 · xhigh: 57.58, $14.7638 per task
- Claude Fable 5.1 · max: 57.27, $18.9197 per task
- GPT-6 Astra · medium: 55.15, $5.8024 per task
- Claude Fable 5.1 · high: 54.85, $12.0769 per task
- Claude Fable 5.1 · medium: 53.94, $8.5845 per task
- Claude Opus 5 · xhigh: 53.33, $18.4431 per task
- Claude Opus 5 · max: 52.12, $18.0882 per task
- GPT-6 Astra · low: 50.91, $4.7191 per task
- GPT-6 Sol · max: 50.3, $4.3979 per task
- Claude Opus 5 · high: 49.7, $14.1281 per task
- Claude Opus 5 · medium: 45.15, $9.6715 per task
- Claude Fable 5 · max: 44.85, $22.0152 per task
- Claude Fable 5.1 · low: 43.33, $7.1476 per task
- GLM-5.3 · max: 41.21, $8.2655 per task
- Grok 4.7 · xhigh: 38.79, $11.1615 per task
- GPT-5.6 Sol · max: 38.48, $7.7021 per task
- GLM-5.3-Flash · none: 35.45, $0.3598 per task
- Claude Opus 5 · low: 34.85, $7.2542 per task
- Qwen3.8 Max · xhigh: 26.97, $14.0475 per task
- Claude Opus 4.8 · max: 23.64, $19.6402 per task
- GPT-5.6 Terra · max: 21.52, $5.2531 per task
- Grok 4.6 · high: 19.7, $10.8836 per task
- Gemini 3.8 Flash · high: 18.18, $5.5417 per task
- GPT-5.6 Luna · max: 17.58, $1.0505 per task
- GPT-6 Luna · max: 15.15, $0.2186 per task
- Muse Spark 1.3 · xhigh: 14.85
- Gemini 3.7 Flash (deprecated) · high: 12.42, $3.8238 per task
- Grok 4.5 · high: 12.42, $6.3458 per task
- Claude Sonnet 5 · max: 12.12, $29.1026 per task

### Terminal-Bench 4.0 (Vals AI run) ⭐⭐

Independent run of Terminal-Bench, but one effort per model and repeated unexplained re-runs and cost rewrites.

Vals AI's own Terminal-Bench 4.0 run in mini-swe-agent (8-hour limit per task), one effort per model.

- Claude Fable 5.2 · high: 65.152, $14.564414 per task
- Claude Opus 5.5 · max: 65.152, $12.042479 per task
- Claude Sonnet 5.5 · max: 64.141, $10.752753 per task
- Kimi K3.1 Pro · max: 63.636, $4.56894 per task
- Gemini 4 Carbon · high: 63.131, $7.452059 per task
- GPT-6 Astra · max: 59.596, $9.584259 per task
- Claude Fable 5.1 · max: 58.081, $17.175337 per task
- Muse Wonder Full · max: 58.081, $10.528007 per task
- Gemini 4 Argon · high: 57.576, $17.63994 per task
- GPT-6.1 Sol · max: 55.051, $1.72496 per task
- Claude Opus 5 · max: 53.535, $18.596913 per task
- GPT-6.1 Terra · max: 49.495, $2.15663 per task
- Muse Wonder · max: 49.495, $14.657694 per task
- GPT-6 Sol · max: 44.444, $5.79322 per task
- Claude Fable 5 · max: 41.414, $30.339756 per task
- Muse Ember · max: 39.394, $0.218381 per task
- HY4 · high: 39.394, $1.280113 per task
- GLM-5.3 · max: 38.889, $9.37436 per task
- Qwen3.8 Max Internal · high: 38.889, $12.146852 per task
- GPT-5.6 Sol · max: 37.879, $7.983597 per task
- GLM-5.3-Flash · max: 37.374, $1.219907 per task
- Claude Haiku 5.5 · max: 36.364, $3.117464 per task
- Qwen3.8 Max · xhigh: 34.343, $10.633329 per task
- DeepSeek V4.1 Flash 1017 · max: 31.818, $0.539359 per task
- Step 5 Preview · high: 31.818, $3.113224 per task
- MiMo-V2.6-Pro · high: 31.313, $0.456609 per task
- Grok 4.7 · xhigh: 28.788, $18.011582 per task
- Mistral Large 4 · high: 27.778, $5.180898 per task
- MiniMax-M3.1-Flash · max: 25.253, $0.734722 per task
- Muse Spark 1.3 · max: 24.747, $6.654565 per task
- MiMo-V2.6-Flash · high: 24.242, $0.211804 per task
- Claude Opus 4.8 · max: 23.232, $17.140717 per task
- GPT-5.6 Terra · max: 22.727, $5.602118 per task
- Ember 1 · high: 20.707, $19.484094 per task
- DeepSeek V4.1 Flash · high: 19.697, $0.499473 per task
- Gemini 3.8 Flash · high: 19.697, $8.64749 per task
- DeepSeek V4 Flash 0731 · high: 18.687, $0.766472 per task
- Grok 4.6 · high: 17.172, $5.062449 per task
- Kimi K3 · max: 17.172, $12.478176 per task
- DeepSeek V4 Pro 0813 · max: 14.141, $3.311817 per task
- GPT-6 Luna · max: 13.636, $0.354726 per task
- Inkling 2 Small · high: 12.626, $8.218043 per task
- Gemini 3.7 Flash (deprecated) · high: 12.121, $8.705677 per task
- GPT-5.6 Luna · max: 11.616, $0.731646 per task
- Claude Sonnet 5 · max: 10.606, $26.04996 per task
- Muse Spark 1.3 · xhigh: 10.606, $7.036467 per task
- Grok 4.5 · high: 8.586, $5.51917 per task
- Muse Spark Open · xhigh: 6.566, $7.264062 per task
- Gemini 3.5 Flash · high: 6.061, $5.981757 per task
- Muse Spark 1.2 · xhigh: 6.061, $4.135537 per task
- Beam 1.0 · high: 4.04
- Mercury 3 · high: 3.03, $0.421803 per task
- Gemini 3.1 Pro (preview) · high: 2.525, $4.111957 per task
- GPT-5.4-mini · xhigh: 2.525, $1.438224 per task
- MiniMax-M3 · high: 1.01, $7.229497 per task
- Inkling · high: 0.505, $2.495998 per task
- MiMo V2.5 Pro · high: 0.505, $0.23303 per task
- Inkling Small · high: 0, $0.321679 per task
- Nemotron 3.5 Super VL 120B A12B · high: 0, $0.862751 per task
- Mercury 2.5 · high: 0, $0.368017 per task

### CursorBench 4.0 ⭐⭐

Real Cursor sessions at every effort, but run by Cursor in its own agent on private tasks, with undated and partly estimated points.

Ambiguous multi-file tasks from real Cursor sessions, run by Cursor in its own agent for every model and effort.

- Gemini 4 Carbon · high: 60.4, $9.22 per task
- Gemini 4 Carbon · medium: 60.1, $8.49 per task
- Claude Fable 5.2 · max: 60, $32.79 per task
- Gemini 4 Carbon · low: 58.6, $11.89 per task
- Claude Opus 5.5 · max: 57.8, $13.43 per task
- Claude Fable 5.2 · high: 56.9, $9.7 per task
- Claude Fable 5.2 · xhigh: 56.3, $15.48 per task
- Claude Opus 5.5 · xhigh: 56, $6.98 per task
- Claude Opus 5.5 · high: 56, $3.97 per task
- Claude Sonnet 5.5 · max: 55.5, $7.05 per task
- Claude Fable 5.2 · medium: 54.8, $4.66 per task
- Claude Fable 5.2 · low: 53.2, $3.94 per task
- Claude Sonnet 5.5 · xhigh: 53.1, $2.81 per task
- Gemini 4 Argon · high: 52.9, $4.57 per task
- Gemini 4 Argon · medium: 52.8, $4.39 per task
- Claude Opus 5.5 · medium: 52.5, $2.91 per task
- Claude Fable 5.1 · max: 51.8, $17.28 per task
- Claude Fable 5.1 · xhigh: 51.6, $13.01 per task
- Gemini 4 Argon · low: 50.6, $4.09 per task
- Muse Wonder · max: 50.4, $5.25 per task
- Claude Fable 5.1 · high: 49.2, $9.08 per task
- Claude Haiku 5.5 · max: 48.4, $1.12 per task
- Muse Wonder · xhigh: 48.2, $4.24 per task
- Claude Sonnet 5.5 · high: 47.8, $1.2 per task
- DeepSeek V4.1 Flash 1017 · max: 47.7, $0.98 per task
- Claude Fable 5.1 · medium: 46.8, $7.05 per task
- Claude Opus 5 · max: 46.6, $11.95 per task
- Grok 4.7 · xhigh: 46.3, $6.01 per task
- DeepSeek V4.1 Flash 1017 · medium: 46.3, $0.8 per task
- Claude Opus 5 · xhigh: 46.1, $11.43 per task
- Muse Wonder · high: 46.1, $3.51 per task
- DeepSeek V4.1 Flash 1017 · high: 45.6, $0.77 per task
- Muse Wonder · medium: 45.2, $2.82 per task
- Claude Fable 5.1 · low: 45.1, $5.44 per task
- DeepSeek V4.1 Flash 1017 · low: 44.9, $0.6 per task
- Claude Opus 5 · high: 44.7, $9 per task
- Claude Haiku 5.5 · xhigh: 44.3, $0.56 per task
- Grok 4.7 · high: 43.9, $4.69 per task
- Claude Opus 5.5 · low: 43.7, $1.17 per task
- Claude Opus 5 · medium: 43.3, $6.94 per task
- GLM-5.3 · max: 42.6, $5.05 per task
- Claude Haiku 5.5 · high: 42.3, $0.32 per task
- GPT-5.6 Sol · max: 41.7, $8.23 per task
- Muse Wonder · low: 41.7, $1.91 per task
- Grok 4.7 · medium: 41.6, $3.49 per task
- Muse Spark 1.3 · max: 41.6, $2.64 per task
- Grok 4.6 · xhigh: 41.4, $6.1 per task
- GPT-5.6 Terra · max: 41.3, $5.14 per task
- Claude Opus 5 · low: 40.7, $4.87 per task
- Grok 4.6 · high: 40.4, $5.2 per task
- Gemini 3.8 Flash · high: 39.6, $4.7 per task
- Claude Sonnet 5.5 · medium: 39.2, $0.52 per task
- GLM-5.3 · high: 38, $3.24 per task
- GPT-5.6 Sol · xhigh: 37.7, $4.4 per task
- Muse Spark 1.3 · xhigh: 37.5, $2.1 per task
- Gemini 3.8 Flash · medium: 37.3, $4.06 per task
- Claude Haiku 5.5 · medium: 36.9, $0.17 per task
- GLM-5.3-Flash · max: 36.8, $0.39 per task
- Grok 4.6 · medium: 36.1, $3.48 per task
- GPT-5.6 Luna · max: 35.9, $1.03 per task
- Claude Sonnet 5.5 · low: 35.8, $0.37 per task
- GPT-5.6 Sol · high: 35.7, $2.85 per task
- Claude Sonnet 5 · max: 34.1, $7.17 per task
- GPT-5.6 Terra · xhigh: 33.6, $1.81 per task
- Grok 4.6 · low: 33.4, $2.25 per task
- Muse Spark 1.3 · high: 33.4, $1.66 per task
- GLM-5.3 · low: 33.3, $2.04 per task
- Grok 4.7 · low: 33.1, $1.58 per task
- GPT-5.6 Luna · xhigh: 33, $0.44 per task
- Muse Spark 1.3 · medium: 32.6, $1.49 per task
- Claude Sonnet 5 · xhigh: 32, $4.55 per task
- GPT-5.6 Sol · medium: 31.1, $1.77 per task
- GLM-5.3-Flash · high: 31.1, $0.25 per task
- Claude Haiku 5.5 · low: 30.9, $0.08 per task
- Claude Sonnet 5 · high: 30.8, $3.48 per task
- GPT-5.6 Terra · high: 30.7, $1.11 per task
- GPT-5.6 Luna · high: 29.4, $0.25 per task
- Muse Spark 1.3 · low: 29.3, $0.93 per task
- Claude Sonnet 5 · medium: 28, $2.31 per task
- Composer 2.5 · high: 27.7, $0.68 per task
- GPT-5.6 Terra · medium: 27.6, $0.64 per task
- GLM-5.3-Flash · low: 26.9, $0.15 per task
- GPT-5.6 Terra · low: 25.2, $0.52 per task
- GPT-5.6 Sol · low: 24.6, $0.87 per task
- Claude Sonnet 5 · low: 24.1, $1.39 per task
- GPT-5.6 Luna · medium: 22.2, $0.08 per task
- GPT-5.6 Luna · low: 16, $0.03 per task

### Vibe Code Bench v1.1 (Vals AI) ⭐⭐

Independent and close to vibe coding, but narrow (complete apps from a product brief) and one effort per model.

Building complete applications from a product description — the closest benchmark to vibe coding. One effort per model.

- Claude Fable 5.2 · high: 93.771, $22.13226 per task
- Claude Sonnet 5.5 · max: 92.392, $19.729112 per task
- Gemini 4 Argon · high: 91.906, $8.10797 per task
- Claude Haiku 5.5 · max: 90.444, $6.071856 per task
- Claude Fable 5 · max: 90.352, $41.714694 per task
- Claude Opus 5.5 · max: 90.294, $35.774773 per task
- Claude Fable 5.1 · max: 90.263, $33.367507 per task
- GPT-6 Astra · max: 89.594, $38.512323 per task
- Gemini 4 Carbon · high: 89.06, $8.254926 per task
- GPT-6.1 Sol · max: 88.931, $6.234794 per task
- Claude Opus 5 · high: 88.404, $33.876235 per task
- GPT-6 Sol · max: 87.824, $26.356012 per task
- DeepSeek V4.1 Flash 1017 · max: 87.73, $0.579984 per task
- Grok 4.7 · xhigh: 86.175, $15.82706 per task
- Muse Wonder Full · max: 85.996, $2.293448 per task
- Muse Spark 1.3 · max: 85.856, $2.543084 per task
- MiMo-V2.6-Pro · high: 85.223, $1.043846 per task
- Kimi K3 · max: 84.968, $10.007576 per task
- Kimi K3.1 Pro · max: 84.894, $7.485151 per task
- DeepSeek V4.1 Flash · high: 84.739, $0.407445 per task
- GPT-6.1 Terra · max: 83.997, $5.996953 per task
- Ember 1 · high: 83.869, $8.559507 per task
- Muse Ember · max: 83.703, $0.054457 per task
- Muse Wonder · max: 83.089, $1.802723 per task
- Muse Spark 1.3 · xhigh: 82.859, $2.101762 per task
- Claude Opus 4.8 · max: 82.725, $26.879989 per task
- DeepSeek V4 Pro 0813 · max: 82.297, $0.356023 per task
- GPT-6 Luna · max: 81.649, $1.345703 per task
- Mistral Large 4 · high: 81.555, $11.723179 per task
- Inkling 2 Small · high: 81.546, $1.901278 per task
- Claude Sonnet 5 · max: 81.325, $25.38542 per task
- HY4 · high: 81.008, $2.394585 per task
- GPT-5.6 Sol · max: 80.495, $33.403634 per task
- Muse Spark 1.2 · xhigh: 79.1, $1.529642 per task
- MiMo-V2.6-Flash · high: 78.96, $0.563041 per task
- Gemini 3.8 Flash · high: 78.651, $6.865141 per task
- GLM-5.3 · max: 78.125, $12.450185 per task
- Claude Opus 4.8 · high: 77.485, $6.8557 per task
- Qwen3.8 Max Internal · high: 77.141, $11.668329 per task
- GPT-5.6 Luna · max: 77.055, $0.726426 per task
- Grok 4.6 · high: 76.239, $4.875513 per task
- Muse Spark Open · xhigh: 76.187, $1.655609 per task
- DeepSeek V4 Flash 0731 · high: 74.736, $0.905592 per task
- GPT-5.6 Terra · max: 74.594, $7.887015 per task
- MiniMax-M3.1-Flash · max: 73.944, $1.146649 per task
- Muse Spark 1.1 · xhigh: 72.158, $0.923283 per task
- Claude Opus 4.7 · max: 71.003, $21.407199 per task
- Step 5 Preview · high: 70.399, $5.856257 per task
- Gemini 3.7 Flash (deprecated) · high: 70.395, $4.827643 per task
- GPT-5.5 · xhigh: 69.847, $16.661911 per task
- Grok 4.5 · high: 69.002, $4.102194 per task
- GPT-5.4 · xhigh: 67.421, $10.686976 per task
- GPT-5.5 Factory · high: 67.391, $5.9268 per task
- Beam 1.0 · high: 66.436
- Qwen3.8 27B · xhigh: 64.85, $23.456372 per task
- Qwen3.8 Max · xhigh: 64.696, $8.238852 per task
- Gemini 3.6 Flash · high: 64.005, $3.041993 per task
- GLM-5.2 · max: 63.961, $8.488514 per task
- GPT-5.3 Codex · xhigh: 61.767, $11.914536 per task
- Mercury 3 · high: 60.927, $0.627191 per task
- GPT-5.5 Codex · high: 58.194, $3.6264 per task
- Claude Opus 4.6 · max: 57.573, $8.687436 per task
- Claude Sonnet 4.6 · high: 55.774, $7.354471 per task
- GPT-5.2 · xhigh: 53.499, $17.745182 per task
- Claude Sonnet 4.6 · max: 51.476, $5.907892 per task
- Composer 2.5 · high: 49.614, $0.502233 per task
- Gemini 3.5 Flash · high: 48.683, $2.540385 per task
- GPT-5.4 · high: 48.474, $7.50518 per task
- GPT-5.4-mini · xhigh: 47.969, $1.191299 per task
- Qwen3.7 Max · max: 47.671, $11.400322 per task
- MiniMax-M3 · high: 47.566, $6.453899 per task
- Kimi K2.7 Code · high: 47.211, $3.835806 per task
- Qwen3.7 Plus · high: 46.391, $1.082348 per task
- MiMo V2.5 · high: 42.174, $0.068395 per task
- GPT-5.2 Codex · high: 37.912, $4.152495 per task
- Kimi K2.6 · high: 37.891, $1.925351 per task
- Gemini 3.5 Flash-Lite · high: 37.161, $0.833111 per task
- MiMo V2.5 Pro · high: 34.113, $0.166747 per task
- Gemini 3.1 Pro (preview) · high: 32.034, $3.825202 per task
- GLM-5.1 · high: 31.456, $2.894911 per task
- GLM-5.3-Flash · max: 30.759, $4.69761 per task
- GPT-5.4 Nano · high: 26.097, $1.277298 per task
- Qwen3.6 Plus · high: 25.565, $5.447636 per task
- GPT-5.1 · high: 24.606, $2.574926 per task
- GLM-5 · high: 23.359, $40.270258 per task
- Claude Sonnet 4.5 · high: 22.621, $6.656984 per task
- GPT-5.1 Codex · high: 22.168, $7.324105 per task
- Claude Opus 4.5 · high: 20.63, $32.873283 per task
- Gemini 3 Flash · high: 20.204, $0.942067 per task
- GPT-5 · high: 20.088, $1.525167 per task
- Muse Spark · high: 19.674, $0.579705 per task
- Grok 4.3 · high: 19.403, $1.28472 per task
- Inkling · high: 19.214, $1.605949 per task
- Inkling Small · high: 19.052, $0.432264 per task
- Kimi K2.5 · high: 17.536, $0.879292 per task
- Qwen3.5 Plus · high: 15.738, $3.803761 per task
- MiniMax-M2.5 · high: 14.853, $2.197729 per task
- Gemini 3 Pro · high: 14.3, $6.417592 per task
- GPT-5 Mini · high: 14.171, $0.248587 per task
- Grok Build 0.1 · high: 13.349
- Qwen3.6 27B · high: 11.941, $4.655127 per task
- MiniMax-M2.7 · high: 11.926, $2.819109 per task
- Claude Haiku 4.5 · high: 11.393, $1.306474 per task
- Laguna M.1 · high: 11.038, $2.059401 per task
- Nemotron 3 Ultra 550B A55B · high: 7.636
- SWE 1.6 Fast · high: 7.22
- Laguna Xs.2 · high: 5.212, $0.981155 per task
- DeepSeek V3p2 · high: 5.108, $2.469325 per task
- Grok 4.20 0309 Reasoning · high: 4.063, $0.771885 per task
- Nemotron 3.5 Super VL 120B A12B · high: 3.71, $0.539233 per task
- Mercury 2.5 · high: 3.683, $0.378016 per task
- Qwen3 Max · max: 3.506, $6.023105 per task
- Ling 3.0 Flash Af Rc3 · high: 3.4, $0.233022 per task
- GLM-4.6 · high: 3.09, $10.845475 per task
- Ling 3.0 Flash 2607 · high: 2.914, $0.292199 per task
- Mistral Medium 3.5 · high: 2.888, $16.735503 per task
- Grok 4.1 Fast Reasoning · high: 1.2, $0.209444 per task
- Gemini 2.5 Pro · high: 0.4, $1.223831 per task
- Gemini 3.1 Flash Lite · high: 0, $0.51 per task
- Nemotron Lightning 3p5 30B A3B · high: 0, $0.128799 per task
- Grok 4 Fast Reasoning · high: 0, $0.133206 per task
- Mistral Small 2603 · high: 0, $6.473001 per task

### DeepSWE v1.1 ⭐⭐

Independent long software-engineering tasks, but updated less often than the others: the newest models are often missing.

Long software-engineering tasks run by Datacurve with the mini-swe-agent harness (pass@1 over 4 runs).

- GPT-6 Astra · xhigh: 74.115, $4.4291 per task
- Gemini 3.8 Flash · high: 73.8255, $2.3623 per task
- Claude Opus 5 · max: 73.6486, $11.8376 per task
- GPT-6 Astra · high: 73.2301, $3.9237 per task
- GPT-6 Astra · max: 73.2301, $7.4978 per task
- Claude Opus 5 · xhigh: 73.1544, $9.0722 per task
- Claude Opus 5 · high: 72.8285, $6.0761 per task
- GPT-6 Astra · medium: 72.7876, $3.0755 per task
- GPT-5.6 Sol · max: 72.6667, $6.456 per task
- Gemini 3.8 Flash · medium: 71.0177, $1.9671 per task
- GPT-5.6 Sol · xhigh: 70.7317, $3.5986 per task
- Claude Fable 5 · xhigh: 69.9115, $13.4145 per task
- Claude Fable 5 · max: 69.7248, $21.6347 per task
- GPT-5.6 Terra · max: 69.6231, $3.9567 per task
- GPT-5.6 Sol · high: 69.4013, $2.6618 per task
- GLM-5.3 · max: 68.9579, $3.9934 per task
- Claude Opus 5 · medium: 68.9038, $3.2898 per task
- Claude Fable 5 · high: 68.6047, $9.1776 per task
- Kimi K3 · max: 68.5144, $4.6547 per task
- Grok 4.6 · medium: 67.4779, $3.449 per task
- GPT-5.6 Luna · max: 67.1875, $0.6056 per task
- GPT-5.5 · xhigh: 67.0354, $7.2262 per task
- GPT-6 Astra · low: 67.0354, $1.5952 per task
- Grok 4.6 · xhigh: 66.7406, $5.4977 per task
- Gemini 3.7 Flash (deprecated) · medium: 65.4867, $2.0251 per task
- Claude Fable 5 · medium: 65.367, $6.0882 per task
- Gemini 3.7 Flash (deprecated) · high: 65.2655, $2.1763 per task
- Grok 4.6 · high: 65.1885, $4.3849 per task
- GPT-5.5 · high: 64.3805, $5.1004 per task
- GLM-5.3-Flash · max: 63.3929, $0.241 per task
- DeepSeek V4 Pro 0813 · max: 62.8319, $1.666 per task
- GPT-5.6 Sol · medium: 61.0619, $1.4158 per task
- GPT-5.6 Terra · xhigh: 60.177, $1.7018 per task
- Claude Fable 5 · low: 59.5843, $3.7579 per task
- Claude Opus 4.8 · max: 58.9744, $13.2226 per task
- Claude Opus 5 · low: 58.1292, $1.6626 per task
- Qwen3.8 Max · xhigh: 57.461, $3.7291 per task
- GPT-5.6 Luna · xhigh: 56.8584, $0.3071 per task
- Muse Spark 1.2 · xhigh: 54.8673, $3.6955 per task
- Claude Opus 4.8 · xhigh: 54.3624, $8.0064 per task
- GPT-5.5 · medium: 53.9823, $2.7493 per task
- Claude Sonnet 5 · max: 53.8462, $26.3999 per task
- Gemini 3.7 Flash (deprecated) · low: 53.7611, $1.8323 per task
- GPT-5.6 Terra · high: 53.7611, $0.9075 per task
- Grok 4.5 · high: 53.7611, $2.4157 per task
- DeepSeek V4 Flash 0731 · max: 53.3186, $0.4641 per task
- Muse Spark 1.1 · xhigh: 53.3186, $2.3611 per task
- Claude Opus 4.8 · high: 51.7699, $4.2819 per task
- GPT-5.4 · xhigh: 51.7699, $5.6525 per task
- Claude Sonnet 5 · xhigh: 49.6674, $11.8906 per task
- Claude Opus 4.8 · medium: 48.6726, $3.4439 per task
- Claude Sonnet 5 · high: 48.2301, $7.4256 per task
- Gemini 3.6 Flash · high: 46.6814, $2.2095 per task
- GPT-5.6 Sol · low: 45.354, $0.817 per task
- GPT-5.6 Luna · high: 44.2478, $0.1556 per task
- GLM-5.2 · max: 43.7778, $3.9199 per task
- Grok 4.6 · low: 41.6481, $1.0424 per task
- Claude Opus 4.8 · low: 40.7982, $2.2934 per task
- Claude Sonnet 5 · medium: 39.7778, $4.079 per task
- GLM-5.2 · high: 36.2832, $2.8355 per task
- Gemini 3.5 Flash · high: 36.0619, $3.4467 per task
- GPT-5.6 Terra · medium: 35.1111, $0.4666 per task
- Kimi K2.7 Code · high: 30.531, $2.8155 per task
- Claude Sonnet 5 · low: 30.5122, $2.1866 per task
- Claude Sonnet 4.6 · high: 29.9335, $5.5224 per task
- GPT-5.5 · low: 26.9912, $1.2002 per task
- GPT-5.6 Terra · low: 24.0535, $0.3422 per task
- Gemini 3.1 Pro (preview) · high: 11.7257, $2.1434 per task
- GPT-5.6 Luna · medium: 11.2832, $0.0433 per task
- GPT-5.6 Luna · low: 1.5487, $0.0145 per task

### Artificial Analysis Intelligence Index v5 ⭐

General intelligence index, not coding-only: useful context, not a coding ranking.

General index of nine evaluations (agentic knowledge work, Terminal-Bench 4.0, FrontierCode 2.0, Terminal-Bench-Science, HLE-Diamond, CritPt…), not coding-only; the broadest per-effort coverage available.

- Claude Fable 5.2 · max: 64.5636, $26.5386 per task
- Claude Fable 5.2 · xhigh: 62.2167, $14.8286 per task
- Claude Fable 5.2 · high: 61.8219, $9.0784 per task
- Gemini 4 Carbon · high: 60.0257, $7.7555 per task
- Claude Opus 5.5 · max: 59.106, $10.0364 per task
- Gemini 4 Carbon · medium: 58.2588, $6.5858 per task
- Claude Opus 5.5 · xhigh: 57.4447, $6.0472 per task
- GPT-6 Astra · max: 56.8539, $6.2447 per task
- Claude Fable 5.2 · medium: 56.4083, $3.8026 per task
- GPT-6 Astra · xhigh: 55.3057, $3.6589 per task
- GPT-6 Astra · high: 54.8498, $3.2386 per task
- GPT-6.1 Sol · max: 54.6087, $1.2721 per task
- Claude Sonnet 5.5 · max: 54.5413, $10.0206 per task
- Gemini 4 Carbon · low: 54.2206, $7.1866 per task
- Claude Fable 5.2 · low: 53.9249, $3.0972 per task
- GPT-6.1 Sol · xhigh: 52.4582, $0.764 per task
- Claude Opus 5.5 · high: 51.836, $3.4647 per task
- GPT-6 Astra · medium: 51.5548, $2.7324 per task
- GPT-6.1 Terra · max: 51.3265, $0.9611 per task
- Gemini 4 Argon · high: 50.4122, $9.9466 per task
- GPT-6.1 Sol · high: 50.3291, $0.591 per task
- Muse Wonder Full · max: 50.0772
- Claude Fable 5.1 · xhigh: 50.0216, $10.8475 per task
- Claude Sonnet 5.5 · xhigh: 49.924, $4.092 per task
- Claude Fable 5.1 · max: 49.1602, $12.994 per task
- Claude Opus 5.5 · medium: 48.2081, $2.4511 per task
- Claude Fable 5.1 · high: 47.175, $7.54 per task
- Kimi K3.1 Pro · max: 46.4526, $3.6812 per task
- GPT-6.1 Sol · medium: 46.3469, $0.3594 per task
- GPT-6.1 Terra · xhigh: 45.2501, $0.4007 per task
- Claude Opus 5 · max: 44.1646, $11.8215 per task
- GPT-6 Sol · max: 44.1555, $2.4658 per task
- Muse Wonder · max: 43.5748, $6.952 per task
- GPT-6 Astra · low: 43.335, $1.4237 per task
- Claude Fable 5.1 · medium: 43.1716, $5.5643 per task
- GPT-6.1 Terra · high: 42.4345, $0.2593 per task
- Claude Sonnet 5.5 · high: 41.5151, $1.6361 per task
- Muse Wonder · high: 40.5266, $4.5605 per task
- Claude Fable 5.1 · low: 38.6836, $4.4711 per task
- GPT-6.1 Terra · medium: 38.1652, $0.1688 per task
- Muse Spark 1.3 · max: 36.6172, $2.3858 per task
- Claude Haiku 5.5 · max: 36.4845, $0.411 per task
- Muse Ember · max: 36.0213, $0.1359 per task
- Grok 4.7 · xhigh: 35.485, $7.7715 per task
- Grok 4.7 · high: 35.2123, $6.1836 per task
- GPT-6.1 Sol · low: 34.8029, $0.1615 per task
- Claude Opus 5.5 · low: 34.7528, $0.9755 per task
- Claude Haiku 5.5 · xhigh: 34.6388, $0.2484 per task
- Tencent Hy4 · high: 34.5552, $1.0311 per task
- Muse Ember · xhigh: 34.115, $0.1162 per task
- Qwen3.8 Max · xhigh: 33.6546, $10.0239 per task
- DeepSeek V4.1 Flash 1017 · max: 32.5896, $0.6021 per task
- GPT-5.6 Terra · max: 32.3434, $3.7349 per task
- GLM-5.3 · max: 31.6534, $5.0093 per task
- Grok 4.7 · low: 31.1321, $2.9222 per task
- Grok 4.6 · xhigh: 31.0288, $4.6786 per task
- Gemini 3.8 Flash · high: 30.8468, $7.766 per task
- Claude Sonnet 5.5 · medium: 30.8196, $0.7807 per task
- Inkling 2 Small · high: 30.6287, $1.1784 per task
- GPT-6.1 Terra · none: 30.5159, $0.2259 per task
- Gemini 3.8 Flash · medium: 30.0786, $6.2586 per task
- MiMo-V2.6-Pro · high: 29.8898, $0.2572 per task
- Claude Haiku 5.5 · high: 29.8743, $0.1586 per task
- Ling 3.1 Flash 1019 · high: 29.5701, $3.1851 per task
- GPT-6.1 Terra · low: 29.4169, $0.0713 per task
- MiniMax-M3.1-Flash · max: 29.0595, $0.491 per task
- DeepSeek V4.1 Flash 1017 · high: 28.6437, $0.4562 per task
- Step 5 Preview · high: 28.5283, $2.3453 per task
- GPT-6 Luna · max: 27.8552, $0.152 per task
- GLM-5.3-Flash · max: 27.8039, $0.5833 per task
- MiMo-V2.6-Flash · high: 26.7277, $0.1285 per task
- Mistral Large 4 · high: 26.125, $2.0617 per task
- Claude Sonnet 5.5 · low: 25.6497, $0.5113 per task
- Muse Spark Open · xhigh: 24.8816, $2.0177 per task
- Claude Haiku 5.5 · medium: 24.6013, $0.0918 per task
- Agnes 2.5 Pro Theta · high: 24.029, $0.5747 per task
- GPT-5.6 Luna · max: 23.7902, $0.3313 per task
- GPT-6 Luna · xhigh: 23.4509, $0.1041 per task
- Mistral Large 4 · medium: 23.1277, $1.4121 per task
- GLM-5.3 · low: 23.1079, $1.5749 per task
- DeepSeek V4 Pro 0813 · max: 23.0633, $1.7691 per task
- Gemini 3.8 Flash · low: 22.5483
- GPT-6 Luna · high: 21.429, $0.0729 per task
- Nanbeige4.3-33B · high: 20.4071
- Qwen3.8 27B · xhigh: 20.1069, $3.5295 per task
- MiniMax-M3 · high: 18.8789, $1.2851 per task
- Claude Haiku 5.5 · low: 18.618, $0.0438 per task
- DeepSeek V4.1 Flash 1017 · none: 18.5132, $0.4254 per task
- GPT-6 Luna · medium: 18.4542, $0.0422 per task
- K2 Horizon 375B A23B · high: 17.8203
- Beam 1.0 · high: 17.7307, $0.2844 per task
- Qwen3.8 27B · medium: 17.4832, $2.16 per task
- Kimi K3 · low: 17.4192, $2.3688 per task
- Qwen3.8 27B · low: 15.6851, $1.9059 per task
- Apodex 1.1 · high: 13.6713, $0.7642 per task
- Nemotron 3.5 Super VL · high: 13.6203, $0.5196 per task
- Inkling Small · high: 12.9102, $0.2032 per task
- DeepSeek V4 Pro 0813 · none: 12.5676, $0.977 per task
- Inkling · xhigh: 12.4782, $1.352 per task
- Nemotron 3 Ultra 550B A55B · high: 11.7593, $2.2257 per task
- Gemini 3.5 Flash-Lite · high: 11.6563, $0.6929 per task
- Qwen3.8 27B · none: 10.0317, $3.7729 per task
- GPT-6 Luna · low: 9.6661, $0.0051 per task
- GPT-6 Luna · none: 9.4424, $0.0714 per task
- K-EXAONE 2.0 · high: 7.5636
- Claude Haiku 4.5 · high: 7.4303, $0.5318 per task
- Muse Glimmer · high: 6.3413, $0.1695 per task
- Command A+ · high: 6.2412
- Nemotron 3.5 Lightning · high: 5.5347, $0.3145 per task
- Granite 4.2 8B · high: 5.0983, $0.0405 per task
- Granite 4.2 3B · high: 4.7227
- Mistral Medium 3.5 · high: 4.6451, $1.0391 per task
- Gemma 4 31B · high: 4.3844
- Mistral Small 4 · high: 4.2972, $0.0399 per task
- GPT-OSS 120B · high: 3.8198, $0.3453 per task
- North Mini Code · high: 3.4247
- NVIDIA Nemotron 3 Nano 4B · high: 1.8158

### Vals Index v2.2 ⭐

Aggregate of professional tasks (finance, tax, legal…), not coding-only.

Vals AI's aggregate of agentic and professional tasks. Version 2.2 (2026-10-19) moves Finance Agent to v2.1, EMB to v1.1, Code Migration to v1.1 and HLAB to HLAB-Vals-1.1, and scores Tax Agent Bench on all-pass accuracy: its scores cannot be compared with v2.1. One effort per model.

- Claude Fable 5.2 · high: 71.309, $17.939202 per task
- Claude Sonnet 5.5 · max: 70.854, $16.183857 per task
- Gemini 4 Carbon · high: 70.45, $10.508065 per task
- Claude Opus 5.5 · max: 70.143, $24.772748 per task
- Gemini 4 Argon · high: 69.389, $13.372731 per task
- Muse Wonder Full · max: 66.787, $6.809627 per task
- GPT-6 Astra · max: 65.023, $21.277828 per task
- Muse Wonder · max: 64.877, $10.647504 per task
- GPT-6.1 Sol · max: 63.944, $3.614054 per task
- Kimi K3.1 Pro · max: 62.52, $6.440192 per task
- DeepSeek V4.1 Flash 1017 · max: 61.257, $0.65747 per task
- GPT-6.1 Terra · max: 60.536, $2.939376 per task
- Muse Ember · max: 60.475, $0.15305 per task
- GLM-5.3 · max: 59.847, $8.524625 per task
- HY4 · high: 59.353, $1.720395 per task
- Qwen3.8 Max Internal · high: 59.045, $8.729516 per task
- Claude Haiku 5.5 · max: 58.727, $3.17475 per task
- Grok 4.7 · xhigh: 58.615, $11.625468 per task
- MiMo-V2.6-Pro · high: 57.813, $0.42958 per task
- Mistral Large 4 · high: 56.527, $3.716552 per task
- Inkling 2 Small · high: 56.081, $3.509318 per task
- Gemini 3.8 Flash · high: 55.951, $5.025617 per task
- MiMo-V2.6-Flash · high: 54.826, $0.221007 per task
- Step 5 Preview · high: 54.762, $2.73409 per task
- MiniMax-M3.1-Flash · max: 53.781, $0.659684 per task
- DeepSeek V4 Pro 0813 · max: 52.974, $2.499691 per task
- Kimi K3 · max: 52.955, $7.169465 per task
- GPT-6 Luna · max: 52.322, $0.450147 per task
- GLM-5.3-Flash · max: 50.861, $2.048666 per task
- Muse Spark Open · xhigh: 48.567, $4.71792 per task
- Mercury 3 · high: 43.659, $0.477863 per task
- Beam 1.0 · high: 43.424
- Nemotron 3.5 Super VL 120B A12B · high: 29.784, $0.627546 per task
- Inkling · high: 28.961, $1.847767 per task
- Inkling Small · high: 27.588, $0.401307 per task

### Terminal-Bench 4.0 (vendor launch charts) ⭐

Figures published by the vendors themselves at launch, each in its own agent: not independent.

Official per-effort curves from Anthropic's Opus 5.5, Sonnet 5.5 and Haiku 5.5 posts (Fable 5.2: max only, from its system card, no cost published), OpenAI's GPT-6 Astra and GPT-6.1 Terra posts (the Terra post also charts GPT-6.1 Sol) and Moonshot's Kimi K3.1 Pro post (Claude Code harness, high and max only), each vendor with its own agent unless stated. Claude costs use the prices in force on 2026-10-07 (Sonnet 5.5 cache reads at $0.10). Compare with the independent runs.

- Claude Sonnet 5.5 · max: 70.6, $10.44 per task
- Claude Fable 5.2 · max: 67.6
- Claude Opus 5.5 · xhigh: 66.4, $7.35 per task
- Kimi K3.1 Pro · max: 65.7, $5.25 per task
- Claude Opus 5.5 · max: 64.8, $11.24 per task
- Claude Opus 5.5 · high: 64.2, $3.88 per task
- Claude Sonnet 5.5 · xhigh: 61.5, $4.34 per task
- GPT-6 Astra · high: 57.88, $7.21 per task
- Claude Opus 5.5 · medium: 57.6, $2.94 per task
- GPT-6 Astra · xhigh: 57.58, $7.48 per task
- GPT-6 Astra · max: 56.67, $10.35 per task
- GPT-6.1 Sol · max: 55.9, $1.98 per task
- Claude Fable 5.1 · max: 55.8, $19.5 per task
- GPT-6 Astra · medium: 53.94, $6.15 per task
- GPT-6.1 Sol · xhigh: 53.4, $1.28 per task
- GPT-6.1 Sol · high: 53, $1.28 per task
- Kimi K3.1 Pro · high: 52.8, $4.16 per task
- GPT-6.1 Sol · medium: 52.7, $1.02 per task
- Claude Fable 5.1 · xhigh: 51.3, $15.8 per task
- GPT-6.1 Terra · max: 50.3, $1.84 per task
- GPT-6 Astra · low: 49.7, $4.95 per task
- Claude Fable 5.1 · high: 49.4, $10.5 per task
- GPT-6.1 Sol · low: 47.1, $0.76 per task
- GPT-6.1 Terra · xhigh: 46.4, $0.99 per task
- Claude Fable 5.1 · medium: 43.4, $7.8 per task
- Claude Sonnet 5.5 · high: 43, $1.46 per task
- GPT-6.1 Terra · high: 41.8, $0.65 per task
- Claude Fable 5.1 · low: 40.2, $5.7 per task
- Claude Haiku 5.5 · max: 39.2, $2.64 per task
- Claude Opus 5.5 · low: 38.5, $1.29 per task
- Grok 4.7 · xhigh: 38
- GPT-6.1 Terra · medium: 35.8, $0.46 per task
- Claude Haiku 5.5 · xhigh: 31.5, $1.75 per task
- Claude Sonnet 5.5 · medium: 28.8, $0.68 per task
- Claude Haiku 5.5 · high: 24.8, $1.04 per task
- GPT-6.1 Terra · low: 24.8, $0.24 per task
- Claude Haiku 5.5 · medium: 20.3, $0.68 per task
- Grok 4.6 · high: 20.3
- Claude Sonnet 5.5 · low: 20, $0.62 per task
- Claude Haiku 5.5 · low: 12.7, $0.42 per task

### FrontierCode 1.1 ⭐

Archive replaced by FrontierCode 2.0; the newest models are not in it.

Would the maintainer merge the change? Tasks written by open-source maintainers, run by Cognition with each lab's own agent, at every effort level (main subset, 100 tasks).

- Claude Opus 5.5 · medium: 54.64, $0.8016 per task
- Claude Opus 5.5 · max: 54.43, $6.191 per task
- Claude Opus 5.5 · high: 53.99, $1.0896 per task
- Claude Fable 5 · xhigh: 53.48, $13.0938 per task
- Claude Opus 5 · medium: 53.38, $4.3127 per task
- GPT-6 Astra · max: 53.26, $4.586 per task
- Claude Fable 5 · high: 52.73, $9.4805 per task
- Claude Sonnet 5.5 · xhigh: 52.09, $1.5881 per task
- Claude Fable 5 · max: 51.57, $19.0696 per task
- Claude Opus 5.5 · xhigh: 51.42, $2.2545 per task
- GPT-6 Astra · high: 50.94, $3.0149 per task
- Claude Fable 5.1 · medium: 50.91, $3.2845 per task
- GPT-6 Astra · xhigh: 50.62, $3.2755 per task
- Claude Fable 5.1 · high: 50.34, $5.2741 per task
- Claude Fable 5.1 · max: 50.28, $12.8257 per task
- GPT-6.1 Sol · medium: 50.23, $0.3607 per task
- SWE-2 · max: 50, $1.1761 per task
- Claude Fable 5 · medium: 49.82, $7.1539 per task
- Claude Fable 5.1 · low: 49.82, $2.3849 per task
- Claude Sonnet 5.5 · high: 49.38, $0.4175 per task
- GPT-6.1 Sol · xhigh: 49.29, $0.5599 per task
- GPT-6 Sol · max: 49.27, $2.0651 per task
- GPT-6 Astra · medium: 48.83, $2.4259 per task
- Claude Fable 5.1 · xhigh: 48.73, $9.2738 per task
- GPT-6 Sol · xhigh: 48.45, $1.325 per task
- Claude Opus 5 · max: 48.04, $11.416 per task
- Claude Fable 5 · low: 48.01, $4.9232 per task
- Grok 4.6 · high: 48.01, $2.8758 per task
- Claude Opus 5 · high: 47.99, $7.2397 per task
- GPT-6.1 Sol · high: 47.98, $0.498 per task
- GPT-6 Sol · high: 47.7, $1.0393 per task
- Grok 4.7 · high: 47.59, $6.6472 per task
- GPT-6.1 Sol · max: 47.57, $0.8481 per task
- Gemini 4 Argon · medium: 47.55, $6.7425 per task
- GPT-5.6 Sol · max: 47.49, $5.1898 per task
- Claude Opus 5.5 · low: 47.3, $0.4036 per task
- SWE-2 · high: 47.2, $0.7813 per task
- Gemini 4 Argon · low: 47.2, $7.3973 per task
- GPT-5.6 Sol · xhigh: 46.84, $4.1454 per task
- Claude Opus 4.8 · max: 46.5, $9.6225 per task
- Claude Haiku 5.5 · max: 46.36, $1.3284 per task
- Claude Sonnet 5.5 · max: 46.21, $20.7818 per task
- Grok 4.6 · medium: 46.06, $1.9762 per task
- GPT-6 Sol · medium: 45.92, $0.7659 per task
- Gemini 4 Argon · high: 45.87, $11.1068 per task
- Claude Haiku 5.5 · xhigh: 45.83, $0.6297 per task
- Grok 4.7 · xhigh: 45.59, $8.2891 per task
- Claude Opus 4.8 · xhigh: 45.49, $6.7803 per task
- GPT-6.1 Sol · low: 45.46, $0.2457 per task
- GPT-6 Astra · low: 45.27, $1.7 per task
- GPT-5.6 Sol · high: 45.06, $3.4801 per task
- Kimi K3 · none: 44.17, $3.8167 per task
- DeepSeek V4.1 Flash · high: 43.87, $0.6202 per task
- Claude Opus 5 · xhigh: 43.65, $9.1423 per task
- Gemini 3.7 Flash (deprecated) · medium: 43.59, $1.8227 per task
- SWE-2 · medium: 43.09, $0.3712 per task
- GPT-5.5 · xhigh: 42.96, $4.0348 per task
- Claude Sonnet 5 · xhigh: 42.73, $10.0666 per task
- Grok 4.5 · high: 42.44, $1.3007 per task
- GPT-6 Luna · max: 42.42, $0.1027 per task
- Claude Sonnet 5 · max: 42.36, $17.1164 per task
- Gemini 3.7 Flash (deprecated) · high: 42.23, $2.1184 per task
- SWE-1.7 · none: 41.99, $1.9734 per task
- Claude Opus 5 · low: 41.95, $2.678 per task
- Grok 4.5 · medium: 41.93, $1.2308 per task
- Claude Haiku 5.5 · high: 41.89, $0.256 per task
- Claude Haiku 5.5 · medium: 41.63, $0.13 per task
- GPT-5.6 Terra · max: 41.31, $1.8735 per task
- Gemini 3.8 Flash · medium: 41.19, $2.5954 per task
- Claude Opus 4.8 · high: 41.02, $4.2799 per task
- GPT-5.5 · high: 40.56, $3.1187 per task
- Claude Opus 4.8 · medium: 40.46, $3.7029 per task
- GLM-5.3 · max: 40.14, $16.9089 per task
- Grok 4.6 · low: 40.11, $0.9474 per task
- GPT-5.6 Sol · medium: 39.93, $2.6897 per task
- GPT-5.6 Luna · max: 39.81, $0.3704 per task
- Claude Sonnet 5 · high: 39.35, $6.0977 per task
- GPT-5.6 Luna · xhigh: 38.91, $0.3093 per task
- GPT-5.6 Terra · xhigh: 38.75, $1.3832 per task
- Claude Opus 4.7 · max: 38.54, $9.0876 per task
- Gemini 3.8 Flash · high: 37.96, $2.598 per task
- Grok 4.5 · low: 37.86, $0.8368 per task
- GPT-6 Sol · low: 37.28, $0.4324 per task
- GPT-6 Luna · high: 37.26, $0.0646 per task
- GPT-6 Luna · xhigh: 37.1, $0.07 per task
- Gemini 3.7 Flash (deprecated) · low: 36.91, $1.7538 per task
- GPT-5.6 Terra · high: 36.89, $1.1875 per task
- GPT-5.5 · medium: 36.6, $2.3513 per task
- Claude Sonnet 5.5 · medium: 36.48, $0.2365 per task
- GPT-5.6 Luna · high: 35.89, $0.2321 per task
- GPT-6 Luna · medium: 35.53, $0.0512 per task
- GPT-5.6 Sol · low: 35.44, $1.8904 per task
- Claude Sonnet 5 · medium: 35.2, $3.8149 per task
- Claude Opus 4.7 · high: 35.17, $5.0292 per task
- Claude Opus 4.8 · low: 35.14, $2.6827 per task
- Claude Opus 4.7 · xhigh: 34.91, $6.8688 per task
- Claude Haiku 5.5 · low: 34.76, $0.0596 per task
- Gemini 3.6 Flash · medium: 34.37, $4.037 per task
- GPT-5.6 Terra · medium: 33.88, $0.8189 per task
- Gemini 3.6 Flash · high: 33.69, $4.2131 per task
- GLM-5.3-Flash · max: 31.83, $1.1466 per task
- GPT-5.5 · low: 30.55, $1.6478 per task
- Kimi K2.7 Code · none: 30.06, $3.0057 per task
- Claude Sonnet 5.5 · low: 29.33, $0.194 per task
- Claude Opus 4.7 · medium: 28.91, $3.2194 per task
- Claude Sonnet 5 · low: 28.71, $2.3897 per task
- DeepSeek V4 Pro 0813 · high: 28.55, $1.812 per task
- Claude Opus 4.7 · low: 27.61, $2.4668 per task
- GPT-5.4-mini · xhigh: 27.04, $1.5201 per task
- Claude Opus 4.6 · high: 26.64, $3.9796 per task
- GPT-5.6 Luna · medium: 25.74, $0.1275 per task
- GPT-6 Luna · low: 25.66, $0.0199 per task
- Composer 2.5 · none: 25.64, $3.0875 per task
- GLM-5.2 · none: 24.5, $2.4704 per task
- Claude Sonnet 4.6 · max: 24.31, $2.8979 per task
- GPT-5.6 Terra · low: 24.11, $0.4961 per task
- Claude Opus 4.6 · medium: 23.67, $2.7941 per task
- Claude Sonnet 4.6 · high: 23.47, $2.6993 per task
- GPT-5.4-mini · high: 23.08, $0.9537 per task
- Gemini 3.6 Flash · low: 22.82, $2.2986 per task
- Claude Sonnet 4.6 · medium: 21.1, $2.3012 per task
- GPT-5.4-mini · medium: 20.78, $0.7833 per task
- DeepSeek V4 Flash 0731 · high: 18.8, $1.5337 per task
- Claude Opus 4.6 · low: 18.31, $2.101 per task
- DeepSeek V4 Pro 0813 · none: 17.64, $1.5548 per task
- GPT-5.6 Luna · low: 15.38, $0.0594 per task
- Claude Sonnet 4.6 · low: 15.03, $1.6481 per task
- MiniMax-M3 · none: 14.7, $0.6756 per task
- Inkling · high: 14, $3.6 per task
- Nemotron 3 Ultra · none: 13.61, $1.4655 per task
- Qwen 3.7 Plus · none: 10.2, $0.2415 per task
- SWE-1.6 · none: 9.39, $0.4891 per task
- GPT-5.4-mini · low: 9.25, $0.3642 per task
- Mistral 3.5 · none: 8, $1.3475 per task

### Z.ai Code Bench (vendor) ⭐

A vendor's private benchmark read from a launch chart, not updated since that launch.

Z.ai's private benchmark run in Claude Code, per effort. Read from the launch chart image (±0.2 pt); no cost is published, so it appears only in the effort chart.

- GLM-5.3 · max: 34.5
- GLM-5.3 · high: 31.4
- GLM-5.3-Flash · max: 29
- GLM-5.3-Flash · high: 28
- GLM-5.3 · low: 24.6
- GLM-5.3-Flash · low: 21.4

## Coding leaderboards

Independent rankings of coding agents and web development. Switch each leaderboard on or off; scroll or zoom inside a chart, and click a provider in its legend to hide it.

### Arena Code: WebDev ⭐⭐⭐

The most cited coding arena: a very large number of blind human votes on real web apps, updated continuously.

Blind pairwise human votes on web-development tasks (995,980 votes, cutoff 2026-11-10). (updated 2026-11-10)

1. Claude Fable 5.2 (xHigh): 1835.7616 ±22.44
2. Claude Opus 5.5 (Max): 1814.2817 ±11.73
3. Claude Sonnet 5.5 (xHigh): 1772.5385 ±9.94
4. Gemini 4 Carbon (High): 1768.1627 ±15.32
5. GPT-6 Astra (Max): 1764.4344 ±8.61
6. GPT-6.1 Sol (Max): 1743.0164 ±9.84
7. Claude Fable 5.1 (Max): 1742.0808 ±7.94
8. Muse Wonder Full (Max): 1723.437 ±14.4
9. Claude Sonnet 5.5 (High): 1715.3326 ±8.42
10. Claude Opus 5 (Max): 1690.9745 ±6.1
11. Kimi K3.1 Pro (Max): 1687.663 ±9.76
12. Gemini 4 Argon (High): 1682.0626 ±9.74
13. GPT-6.1 Terra (Max): 1678.503 ±10.67
14. Qwen3.8-Max (0902): 1675.8153 ±6.25
15. GPT-6 Sol (Max): 1672.967 ±8.36
16. Qwen3.8-Max: 1670.3248 ±12.11
17. DeepSeek-V4.1-Flash-20261017 (Max): 1663.6469 ±6.8
18. Kimi K3 (Max): 1657.8872 ±5.7
19. Claude Opus 5 (High): 1656.6094 ±5.27
20. Muse Spark 1.3 (Max): 1655.0895 ±7.03
21. Muse Wonder (Max): 1648.5945 ±9.46
22. Inkling 2 Small: 1646.2293 ±19.5
23. Grok 4.7 (xHigh): 1640.9524 ±7.95
24. MiMo-V2.6-Flash: 1637.6334 ±7.73
25. Hy4 preview: 1631.6387 ±9.78
26. Claude Fable 5 (High): 1625.3423 ±5.98
27. MiMo-V2.6-Pro: 1624.2596 ±7.67
28. Qwen3.8-Flash-Next: 1623.5135 ±7.11
29. GLM-5.3 (Max): 1622.2916 ±6.87
30. DeepSeek-V4.1-Flash-20260910 (Max): 1620.6257 ±11.13
31. Muse Spark 1.3 (xHigh): 1620.3621 ±6.98
32. Grok 4.6 (High): 1618.0027 ±6.94
33. GPT-5.6 Sol (xHigh, Codex Harness): 1615.3706 ±5.73
34. U2 Flash: 1610.5525 ±16.84
35. GLM-5.3-Flash: 1608.8679 ±6.17
36. Muse Ember (Max): 1608.597 ±16.83
37. GLM-5.2 (Max): 1607.7766 ±5.88
38. Claude Haiku 5.5 (High): 1593.8366 ±8.5
39. Qwen3.8-27B: 1590.2337 ±6.07
40. Gemini 3.7 Flash (High): 1588.0875 ±6.47
41. MiniMax-M3.1-Flash: 1584.5473 ±9.63
42. Gemini 3.8 Flash (High): 1582.646 ±6.65
43. DeepSeek-V4-Flash (High): 1581.6476 ±9.4
44. DeepSeek-V4-Pro (High, 20260813): 1581.1543 ±9.53
45. GPT-6 Luna (Max): 1577.4745 ±5.94
46. Step 5 Preview (High): 1573.3668 ±8.12
47. Mistral Large 4: 1564.4547 ±15.23
48. Claude Opus 4.7 (High): 1557.1039 ±5.8
49. Claude Opus 4.8 (High): 1555.4686 ±5.8
50. Claude Opus 4.7: 1555.3406 ±5.65
51. Grok 4.5: 1552.423 ±6.73
52. Claude Opus 4.6 (High): 1546.6297 ±5.52
53. Muse Spark 1.1: 1541.4551 ±7.68
54. Claude Sonnet 5 (High): 1538.9652 ±5.81
55. Gemini 3.6 Flash (High): 1537.9065 ±7.57
56. Claude Opus 4.6: 1536.7199 ±5.39
57. Claude Opus 4.8: 1536.2549 ±5.74
58. Muse Spark 1.2 (xHigh): 1533.5177 ±13.81
59. Muse Spark Open (Max): 1526.3072 ±15
60. Claude Sonnet 4.6: 1522.1531 ±5.19
61. GPT-5.6 Terra (xHigh, Codex Harness): 1520.1081 ±6.09
62. GPT-5.6 Luna (xHigh, Codex Harness): 1519.107 ±6.46
63. Seed2.1 Pro Preview: 1518.981 ±5.62
64. Qwen3.7-Max (20260517): 1515.0092 ±7.41
65. GPT-5.5 (xHigh, Codex Harness): 1512.1185 ±6.01
66. Kimi K2.6: 1509.2355 ±7.28
67. GLM-5.1: 1508.055 ±6.82
68. Hy3: 1505.6849 ±7.9
69. gemini-3.5-flash: 1505.6456 ±15.55
70. Gemini 3.5 Flash (High): 1499.6653 ±7.32
71. Claude Opus 4.5 (High, 32K, 20251101): 1492.9852 ±8.35
72. Gemini 3.5 Flash (Medium): 1489.7989 ±6.7
73. GPT-5.5 (High, Codex Harness): 1487.0771 ±5.95
74. Spark-X2.6: 1484.7148 ±19.26
75. MiniMax-M3: 1483.9074 ±5.42
76. Qwen3.6-Max-Preview: 1482.2236 ±12.59
77. MiMo-V2.5-Pro: 1479.1321 ±5.27
78. Kimi K2.7 Code: 1474.2541 ±9.45
79. Claude Opus 4.5 (20251101): 1468.5575 ±7.27
80. DeepSeek-V4-Pro Preview (High): 1463.2821 ±6.36
81. GPT-5.4 (High, Codex Harness): 1461.9245 ±18.95
82. Qwen3.6-Plus: 1461.3301 ±5.62
83. GPT-5.5 (Codex Harness): 1455.4137 ±5.92
84. DeepSeek-V4-Pro: 1448.3831 ±6.37
85. Gemini 3.1 Pro Preview: 1446.4896 ±5.1
86. GPT-5.4 (Medium, Codex Harness): 1442.0846 ±18.37
87. Gemini 3 Pro: 1440.613 ±8.5
88. Gemini 3 Flash: 1439.4978 ±8.67
89. MiMo-V2.5: 1437.6244 ±6.06
90. Beam 1.0 501B A23B: 1436.5697 ±13.68
91. Kimi K2.5 (Thinking): 1436.2127 ±5.26
92. Gemini 3.5 Flash-Lite: 1435.2191 ±41.2
93. GLM-5: 1435.1683 ±8.21
94. GLM-4.7: 1435.0362 ±12.06
95. MiMo-V2-Pro: 1434.1396 ±8.2
96. DeepSeek-V4-Flash Preview (High): 1430.7306 ±7
97. GPT-5.2: 1419.7989 ±22.81
98. GPT-5 (Medium): 1416.6611 ±15.67
99. Inkling: 1414.0805 ±5.98
100. Inkling-Small: 1407.5054 ±9.17
101. GPT-5.3-Codex (Codex Harness): 1405.7032 ±13.73
102. Kimi K2.5 (Instant): 1405.3668 ±12.12
103. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B (NVFP4): 1401.2795 ±10.77
104. Qwen3.5-397B-A17B: 1400.068 ±5.17
105. GPT-5.4: 1399.3825 ±9.84
106. GLM-5V-Turbo: 1398.945 ±13.3
107. MiniMax-M2.7: 1397.7833 ±5.84
108. GPT-5.4 mini (High): 1397.7215 ±6.57
109. GPT-5.1 (Medium): 1397.3807 ±11.62
110. Claude Sonnet 4.5 (High, 32K, 20250929): 1393.4648 ±7.78
111. Claude Opus 4.1 (20250805): 1389.4433 ±10.79
112. MiniMax-M2.5: 1385.6231 ±8.39
113. Claude Sonnet 4.5 (20250929): 1384.7493 ±7.34
114. MiniMax-M2.1 Preview: 1384.4482 ±9.6
115. Gemini 3 Flash (Minimal): 1384.3058 ±5.27
116. Grok 4.20 Beta (Reasoning, 0309): 1375.3674 ±6.12
117. GPT-5.3-Codex (Codex Harness): 1371.043 ±11.36
118. Gemma 4 31B: 1369.8572 ±6.01
119. Solar Pro 4: 1369.1087 ±10.48
120. DeepSeek-V3.2 (Thinking): 1362.4619 ±9.16
121. Qwen3.5-122B-A10B: 1360.0422 ±7.7
122. Gemma 4 26B A4B: 1358.4685 ±16.96
123. Hy3 preview: 1358.103 ±17.5
124. Grok 4.3: 1357.3728 ±6.29
125. Qwen3.5-27B: 1357.2303 ±8.08
126. Muse Glimmer: 1354.6515 ±15.7
127. Laguna M.1: 1349.5658 ±9.35
128. GLM-4.6: 1342.2434 ±10.87
129. GPT-5.1: 1342.1709 ±8.55
130. GPT-5.2-Codex: 1338.4514 ±8.99
131. GPT-5.1-Codex: 1338.083 ±11.67
132. MiMo-V2-Flash (Non-Thinking): 1331.0298 ±9.81
133. Claude Haiku 4.5 (20251001): 1330.238 ±4.99
134. DeepSeek-V3.2: 1325.6371 ±8.02
135. Kimi K2 Thinking Turbo: 1321.5181 ±7.51
136. Laguna XS.2: 1303.1774 ±10.74
137. MiniMax-M2: 1299.0726 ±11.02
138. MiMo-V2-Flash (Thinking): 1291.8562 ±16.51
139. Qwen3-Coder-480B-A35B-Instruct: 1274.5225 ±7.86
140. DeepSeek-V3.2-Exp: 1270.407 ±13.62
141. Mistral Medium 3.5: 1263.0425 ±14.87
142. KAT-Coder-Pro V1: 1257.6525 ±19.57
143. Gemini 3.1 Flash-Lite Preview: 1255.5567 ±6.89
144. Qwen3.5-35B-A3B: 1252.8204 ±17.48
145. GPT-5.1-Codex-Mini: 1245.0243 ±22.32
146. Qwen3.5-Flash: 1243.536 ±19.7
147. Grok 4.1 Fast (Reasoning): 1241.8101 ±11.17
148. Trinity Large Thinking: 1236.4382 ±20.07
149. Mistral Large 3: 1233.2878 ±25.9
150. Gemini 2.5 Pro: 1227.5433 ±15.94
151. Grok 4.1 (Thinking): 1216.7135 ±24.94
152. Devstral 2: 1195.5313 ±20.82
153. Granite 4.1 8B: 1192.557 ±18.59
154. Mercury 2: 1170.768 ±24.78
155. Grok Code Fast 1: 1163.9665 ±28.71
156. Grok 4 Fast (Reasoning): 1162.1558 ±27.79
157. Devstral Medium (2507): 1071.8895 ±30.78

### Artificial Analysis Coding Agent Index ⭐⭐⭐

Independent, measures the agents themselves (Claude Code, Codex, Antigravity…) with their models, on current versions.

Each coding agent with its vendor's model (Claude Code, Codex, Grok Build, Kimi Code, Antigravity, OpenCode, Muse Code, Devin Fusion, GitHub Copilot, Mistral Vibe, Mirror, DeepSeek Harness) on FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA — the closest proxy for what a subscription gives you. (updated 2026-11-10)

1. Claude Code - Claude Opus 5.5 (Max, Default Fallback): 61.6529
2. Devin Fusion CLI - Claude Opus 5.5 Max + GPT-6 Luna High: 61.5117
3. GitHub Copilot - Copilot Fusion (max): 59.8571
4. Antigravity CLI - Gemini 4 Carbon (High): 59.4049
5. Claude Code - Claude Sonnet 5.5 (Max, Default Fallback): 59.3374
6. Muse Code - Muse Wonder Full (Max): 58.8804
7. Devin Fusion CLI - Claude Fable 5.1 XHigh + SWE-2 Medium: 58.6619
8. Claude Code - Claude Fable 5.1 (Max, Default Fallback): 57.62
9. Codex - GPT-6 Astra (Max): 54.6636
10. Codex - GPT-6.1 Sol (Max): 54.128
11. Antigravity CLI - Gemini 4 Argon (High): 53.2463
12. Devin Fusion CLI - GPT-6 Astra XHigh + SWE-2 Medium: 53.077
13. Muse Code - Muse Wonder (Max): 51.9768
14. Opencode - GLM-5.3 (Max): 48.9162
15. Codex - GPT-6.1 Terra (Xhigh): 48.2238
16. Codex - GPT-6.1 Terra (Max): 48.1004
17. Codex - DeepSeek V4.1 Flash (Max): 47.3001
18. Grok Build - Grok 4.7 (Xhigh): 46.5969
19. DeepSeek Harness - DeepSeek V4.1 Flash 1017 (Max): 46.3052
20. Codex - GPT-6.1 Terra (High): 46.2788
21. Muse Code - Muse Ember (Max): 45.9458
22. Muse Code - Muse Spark 1.3 (Max): 45.7078
23. Claude Code - Step 5 Preview: 43.1274
24. Codex - GPT-6.1 Terra (Medium): 42.0492
25. Kimi Code CLI - Kimi K3 (Max): 41.4173
26. Muse Code - Muse Ember (Xhigh): 39.8416
27. Claude Code - Qwen3.8 Max: 39.2199
28. Codex - GPT-6.1 Terra (Non-reasoning): 39.2152
29. Claude Code - Claude Haiku 5.5 (Max): 37.3496
30. Mistral Vibe - Mistral Large 4 (High): 35.4072
31. Codex - GPT-6.1 Terra (Low): 34.8568
32. Antigravity SDK - Gemini 3.8 Flash (High): 33.2723
33. Codex - GPT-6 Luna (Max): 31.2156
34. Muse Code - Muse Spark Open (Xhigh): 26.3007
35. Mirror - Beam 1.0: 20.479

### Scale SWE-Bench Pro V2 (hard) ⭐⭐⭐

Independent, real issues of commercial and open-source repositories re-graded on a clean image, current models.

Real issues of commercial and open-source repositories, re-graded on a clean image, each model in its vendor's agent (Claude Code, Codex) or mini-swe-agent. The hard subset: the full set sits near 99% for frontier models. (updated 2026-11-06)

1. Opus 5 (Claude Code) xhigh: 98
2. Fable 5.1 (Claude Code) high: 92.2
3. Gemini 4 Argon (mini-swe-agent) high: 92.16
4. GPT-6 Astra (Codex) high: 90.2
5. Muse Wonder 1.0 (mini-swe-agent) xhigh: 90.2
6. Kimi-K3 (mini-swe-agent) max: 88.2
7. Sonnet 5 (Claude Code) xhigh: 88.2
8. GPT-5.6 Terra (Codex) xhigh: 86.3
9. Inkling 2 Small: 86.27
10. Minimax M3p1 Flash (mini-swe-agent) xhigh: 84.31
11. GLM-5.3 (mini-swe-agent) max: 84.3
12. GPT-5.6 Sol (Codex) xhigh: 82.4
13. Muse Spark Open (mini-swe-agent) xhigh: 82.35
14. Sonnet 5.5 (mini-swe-agent) xhigh: 70.59
15. Gemini 3.8 Flash (mini-swe-agent) high: 58.8
16. Inkling (mini-swe-agent) xhigh: 56.9
17. Haiku 5.5 (mini-swe-agent) high: 52.94 ±13.01
18. Haiku 4.5 (Claude Code) xhigh: 25.5

### LiveBench 2.0: long-horizon ⭐⭐⭐

Independent, contamination-limited long tasks, regularly refreshed, current models.

LiveBench 2.0 (release of 2026-11-06, updated 2026-11-09): 18 long-running agentic tasks in six groups (apps extended over several turns of changing requirements, delegated and long-run operations on a codebase over simulated weeks, systems and performance work), deterministic grading, each model at its highest supported effort unless its name says otherwise, with the cost per task. (updated 2026-11-09)

1. Claude Fable 5.2 (xhigh): 82.9
2. GPT-6 Astra (max): 75.6
3. Claude Fable 5.1 (max): 74.2
4. Claude Opus 5.5 (max): 72.8
5. GPT-6.1 Sol (max): 65.6
6. Claude Sonnet 5.5 (max): 63.4
7. Gemini 4 Argon (high): 58.5
8. GPT-6.1 Terra (max): 56.2
9. Claude Haiku 5.5 (max): 47.9
10. Muse Spark 1.3 (max): 47.9
11. Muse Wonder 1.0 (max): 46.2
12. Grok 4.7 (xhigh): 45.3
13. DeepSeek V4.1 Flash 1017 (max): 38.8
14. Kimi K3.1 Pro (max): 38.7
15. MiMo V2.6 Pro (high): 36.4
16. Kimi K3 (max): 36
17. Qwen3.8 Max (max): 26.6
18. GPT-5.6 Terra (max): 25.9
19. MiniMax M3.1 Flash (max): 25.7
20. GPT-6 Luna (max): 24.7
21. Mistral Large 4 1029 (high): 24.7
22. GLM-5.3 Flash (max): 20.7
23. Step 5 (Preview) (max): 19.4
24. Grok 4.6 (xhigh): 18.7
25. Gemini 3.8 Flash (high): 16.5
26. Qwen3.8 27B (max): 15.9
27. Muse Spark Open (max): 12.1
28. Nemotron 3.5 Super VL (max): 12
29. Smaug-Nemotron-3.5-NVFP4 (max): 11.3
30. Nemotron 3.5 Super VL (NVFP4) (max): 11
31. Inkling (max): 9.8
32. Mistral Large 4 (high): 9.2
33. Gemma 4 31B (max): 8.5

### Arena Agent Arena: code ⭐⭐⭐

Real coding-agent sessions rated by their users, close to a million sessions, current models.

Real coding sessions of Arena users with each model in an agent, judged on signals such as confirmed success, praise and complaints, steering effort and tool errors (909,397 code sessions). Live user sessions, not a fixed task set. (updated 2026-11-08)

1. Claude Opus 5.5 (Max): 0.2292 ±0.031
2. Claude Sonnet 5.5 (Max): 0.1699 ±0.0382
3. Gemini 4 Carbon (High): 0.1557 ±0.0097
4. GPT-6 Astra (Max): 0.143 ±0.0287
5. Claude Fable 5.1 (Max): 0.1426 ±0.0301
6. GPT-6.1 Terra (Max): 0.1398 ±0.0185
7. GPT-6.1 Sol (Max): 0.1252 ±0.0165
8. GPT-6 Sol (Max): 0.1018 ±0.022
9. Muse Wonder Full (Max): 0.0917 ±0.0137
10. Gemini 4 Argon (High): 0.0857 ±0.0118
11. Claude Fable 5 (High): 0.0836 ±0.0166
12. Claude Opus 5 (Max): 0.0769 ±0.0182
13. Kimi K3.1 Pro (Max): 0.0657 ±0.0154
14. Hy4: 0.0504 ±0.0204
15. GPT-5.6 Sol (xHigh): 0.0413 ±0.0148
16. Claude Haiku 5.5 (High): 0.0397 ±0.0139
17. DeepSeek-V4.1-Flash-20261017 (Max): 0.0331 ±0.0075
18. Grok 4.7 (xHigh): 0.0315 ±0.0203
19. Claude Sonnet 5 (High): 0.0297 ±0.0164
20. Muse Spark 1.3 (Max): 0.0268 ±0.0121
21. Muse Wonder (Max): 0.0262 ±0.0179
22. GPT-6 Luna (Max): 0.0247 ±0.0065
23. MiMo-V2.6-Pro: 0.0176 ±0.0124
24. Hy4 preview: 0.0161 ±0.0105
25. GLM-5.2 (Max): 0.0159 ±0.0109
26. GLM-5.3 (Max): 0.0127 ±0.0112
27. Grok 4.6 (xHigh): 0.0074 ±0.0165
28. Inkling 2 Small: 0.0034 ±0.0354
29. MiMo-V2.6-Flash: 0.0005 ±0.0139
30. Kimi K3 (Max): -0.0026 ±0.0081
31. Qwen3.8-Max: -0.0027 ±0.0092
32. GPT-6 Luna (Medium): -0.0067 ±0.0069
33. MiniMax-M3.1-Flash: -0.0099 ±0.0207
34. Gemini 3.8 Flash (High): -0.0142 ±0.0119
35. GPT-5.6 Terra (xHigh): -0.0193 ±0.012
36. Step 5 Preview: -0.0225 ±0.0208
37. GLM-5.3-Flash: -0.0258 ±0.0094
38. Qwen3.8-Flash-Next: -0.0389 ±0.0107
39. Qwen3.8-27B: -0.0408 ±0.0122
40. Gemini 3.7 Flash (High): -0.0469 ±0.0126
41. GPT-6 Luna (Low): -0.0495 ±0.0104
42. Mistral Large 4: -0.0656 ±0.0279
43. Hy3: -0.0679 ±0.0217
44. Beam 1.0 501B A23B: -0.0814 ±0.0374
45. Gemini 3.1 Pro Preview: -0.0828 ±0.0207
46. Qwen3.7-Max: -0.0832 ±0.0162
47. Muse Spark 1.2 (xHigh): -0.0915 ±0.012
48. Spark-X2.6: -0.0928 ±0.0177
49. Muse Spark 1.1: -0.0996 ±0.0099
50. MiniMax-M3: -0.1055 ±0.0156
51. Qwen3.7-Plus: -0.1096 ±0.0214
52. Mistral Medium 3.5: -0.1547 ±0.0245
53. Nemotron 3 Ultra: -0.1555 ±0.0476
54. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B (NVFP4): -0.1666 ±0.0112
55. Inkling: -0.1795 ±0.0227
56. Inkling-Small: -0.2094 ±0.0302
57. Solar Pro 4: -0.259 ±0.031

### FrontierSWE V2 ⭐⭐⭐

Independent, hard software-engineering tasks scored as the mean of five trials, current models.

34 very hard engineering tasks with a 20-hour budget each — the long-horizon extreme. (updated 2026-11-09)

1. Claude Fable 5.2: 69.9402
2. GPT-6 Astra: 65.5076
3. Claude Opus 5.5: 62.331
4. Claude Sonnet 5.5: 61.8893
5. Gemini 4 Carbon: 61.1995
6. Claude Fable 5.1: 56.2908
7. Gemini 4 Argon: 54.9612
8. Claude Opus 5: 52.0118
9. Claude Fable 5: 46.9645
10. Claude Haiku 5.5: 43.7687
11. GPT-5.6 Sol: 32.2025
12. GLM-5.3: 30.1819
13. Grok 4.7: 29.5048
14. Kimi K3: 25.865
15. Muse Spark 1.3: 25.8021
16. Grok 4.6: 25.2862
17. Gemini 3.7 Flash: 20.2606
18. Gemini 3.8 Flash: 19.6269
19. GLM-5.3 Flash: 18.1252
20. Qwen3.8-Max-0902: 17.8096
21. Qwen3.8-Max: 15.8094
22. DeepSeek V4 Flash Vision Exp: 14.8313
23. Muse Spark 1.2: 11.9703
24. Inkling: 4.1189

### SWE-Marathon v1.1 ⭐⭐⭐

Independent long tasks with many trials per model, recent and covering current models.

20 multi-hour engineering tasks, 8 trials each, every model in an agent (mostly Claude Code) at the published effort, with the cost of each trial. (updated 2026-11-07)

1. Antigravity CLI - Gemini 4 Carbon (high): 53.75
2. Muse Code - Muse Wonder Full (max): 53.125
3. Muse Code - Muse Wonder (max): 52.5
4. Claude Code - Claude Sonnet 5.5 (max): 51.25
5. Claude Code - Claude Haiku 5.5 (max): 50
6. Claude Code - Kimi K3 (max): 48.125
7. Muse Code - Muse Spark 1.3 (max): 46.25
8. Claude Code - Claude Opus 5.5 (max): 46.25
9. Claude Code - Claude Fable 5.1 (max): 45.625
10. Grok Build - Grok 4.7 (high): 45.625
11. Claude Code - Claude Opus 5.5 (xhigh): 45.625
12. Gemini CLI - Gemini 4 Argon (high): 43.75
13. Codex - GPT-6-astra (medium): 43.1373
14. Codex - GPT-6-astra (max): 42.5
15. Claude Code - GLM 5.3 (max): 42.5
16. Codex - GPT-6.1 Sol (max): 42.5
17. Codex - GPT-6-astra (low): 41.4474
18. Claude Code - GLM 5.3 (high): 40.625
19. Codex - GPT-6-astra (high): 39.2157
20. Claude Code - Inkling 2 Small (xhigh): 38.75
21. Claude Code - GLM 5.3 Flash (max): 38.75
22. Codex - GPT-6-astra (xhigh): 36.1842
23. Claude Code - GLM 5.3 (low): 29.375
24. Codex - GPT-6 Luna (max): 28.125
25. OpenCode - Gemini 3.7 Flash (max): 22.5
26. Mini-SWE-Agent - DeepSeek V4 Pro (max): 10.625

### LiveBench v1: agentic coding ⭐⭐

Independent and respected, but the v1 suite is now superseded by LiveBench 2.0.

LiveBench v1 (release 2026-06-25, now archived by LiveBench 2.0): contamination-limited questions; agentic coding in JavaScript, TypeScript and Python. (updated 2026-10-16)

1. deepseek-v4.1-flash-max: 77.2727
2. claude-opus-5-5-max-effort: 71.717
3. claude-fable-5-1-max-effort: 66.0607
4. claude-opus-5-5-xhigh-effort: 65.3533
5. claude-opus-5-max-effort: 65.202
6. deepseek-v4-flash-vision-exp: 65.101
7. smaug-agentic: 64.6467
8. qwen3.8-max: 64.6467
9. muse-spark-1.3-xhigh: 64.091
10. muse-wonder-1.0-max: 62.7273
11. claude-fable-5-max-effort: 62.1717
12. kimi-k3: 62.1717
13. qwen3.8-flash-next: 61.6163
14. qwen3.8-27b: 61.3637
15. smaug-flash: 61.0607
16. glm-5.3: 60.909
17. smaug-mini: 60.808
18. claude-sonnet-5-xhigh-effort: 59.394
19. muse-spark-1.1-xhigh: 58.5353
20. gemini-3.7-flash-high: 58.283
21. muse-spark-1.2-xhigh: 57.5757
22. gpt-6-astra-max: 57.323
23. mistral-large-4-high: 57.1717
24. claude-haiku-5-5-max-effort: 57.1717
25. grok-4.6: 57.02
26. gpt-6.1-sol-xhigh: 56.7677
27. glm-5.3-flash: 56.7677
28. grok-4.5: 56.4647
29. claude-sonnet-5-5-max-effort: 56.3133
30. gpt-5.6-sol-max: 56.212
31. gpt-5.6-terra-max: 54.9497
32. deepseek-v4-pro-0813: 54.9497
33. gpt-6.1-sol-max: 54.5453
34. gemini-3.8-flash-high: 54.2423
35. gpt-5.5-xhigh: 53.9897
36. grok-4.7-xhigh: 53.9897
37. gpt-5.4-xhigh: 53.8383
38. gpt-6-sol-max: 52.8787
39. ox-alpha-max: 52.626
40. glm-5.2: 51.7677
41. claude-haiku-5-5-xhigh-effort: 51.3637
42. gpt-6-luna-max: 51.212
43. claude-opus-4-7-xhigh-effort: 50.6563
44. claude-opus-4-8-max-effort: 50.505
45. gpt-5.2-2025-12-11-high: 50.2527
46. gpt-5.2-codex: 49.394
47. inkling-xhigh: 49.394
48. gemini-3.5-flash-high: 48.9897
49. claude-opus-4-6-thinking-auto-high-effort: 48.9897
50. gpt-5.6-luna-max: 48.4343
51. kimi-k2.6-thinking: 46.9193
52. deepseek-v4-flash-0731: 46.7677
53. gpt-5.4-nano-xhigh: 46.7677
54. grok-build-0.1: 45.808
55. kimi-k2.7-code: 45.6563
56. gemini-3.5-flash-lite-high: 45.2527
57. gemini-3.1-pro-preview-high: 44.1413
58. qwen3.7-max: 43.586
59. gemini-3.6-flash-high: 43.4343
60. claude-sonnet-4-6-thinking-auto-medium-effort: 42.626
61. deepseek-v4-pro: 42.626
62. gpt-5.4-mini-xhigh: 41.6667
63. qwen3.6-plus: 41.3637
64. minimax-m3: 40.6563
65. claude-opus-4-5-20251101-thinking-64k-high-effort: 39.697
66. claude-sonnet-5-5-xhigh-effort: 39.3437
67. qwen3.6-27b: 39.2927
68. nemotron-3-ultra-550b-a55b: 38.7373
69. deepseek-v4-flash: 37.626
70. nemotron-3.5-lightning-30b-a3b: 19.9497
71. grok-4.3: 18.5353

### Design Arena: code overall ⭐⭐

Blind human votes on generated code, independent, but fewer votes than Arena and judged on looks first.

All of Design Arena's code categories in one ranking (1,950,799 votes): websites, components, games, SVG, charts and more. No standard error is published, so close Elo scores cannot be told apart. (updated 2026-11-11)

1. GPT-6 Astra (xhigh): 1387
2. Muse Wonder (max): 1383
3. Claude Opus 5.5: 1381
4. Inkling 2 Small: 1379
5. GPT-6 Astra (max): 1370
6. GPT-6.1 Sol: 1354
7. Kimi K3.1 Pro: 1350
8. Kimi K3: 1348
9. Claude Fable 5.2: 1347
10. Muse Spark 1.3 (xhigh): 1346
11. Claude Sonnet 5.5: 1346
12. Muse Spark 1.3 Max: 1345
13. Muse Ember: 1345
14. GPT-5.6 Sol (xhigh): 1335
15. Claude Fable 5.1: 1329
16. Gemini 4 Argon: 1324
17. Claude Opus 5: 1319
18. DeepSeek-V4.1-Flash: 1314
19. MiMo-V2.6-Pro: 1311
20. GPT-5.6 Sol (Medium): 1309
21. Muse Spark 1.2: 1307
22. Claude Fable 5: 1304
23. Gemini 3.7 Flash: 1302
24. GPT-6.1 Terra: 1300
25. Ling-3.1-flash: 1299
26. Muse Spark Open (xhigh): 1298
27. Gemini 3.8 Flash: 1297
28. MiniMax M3.1 Flash: 1293
29. GPT-6 Sol: 1292
30. GLM-5.3: 1290
31. DeepSeek-V4.1-Flash (1017): 1289
32. Qwen3.8 Max: 1288
33. Grok 4.6: 1285
34. Gemini 3.6 Flash: 1284
35. Claude Opus 4.6: 1283
36. GLM 5.2: 1280
37. Claude Opus 4.6 (Thinking): 1279
38. Claude Haiku 5.5: 1278
39. GLM-5.3-Flash: 1274
40. Deca 2.6 Ultra: 1271
41. Claude Sonnet 4.6: 1270
42. Muse Spark 1.1: 1270
43. Claude Sonnet 5: 1269
44. Qwen3.7 Max: 1269
45. Grok 4.5: 1269
46. Grok 4.7: 1268
47. Kimi K2.6: 1265
48. Deca 2.5 Mini: 1265
49. MiMo-V2.5-Pro: 1264
50. GLM 5.1: 1263
51. Qwen3.7 Plus: 1263
52. Gemini 3.5 Flash: 1262
53. GPT-6 Luna: 1262
54. Deca 2.5 Ultra: 1257
55. Hy4: 1256
56. MiMo-V2.5: 1256
57. GLM 5 Turbo: 1255
58. GPT-5.5: 1254
59. Mistral Large 4: 1253
60. Claude Opus 4.8: 1250
61. Kimi K2.7 Code: 1247
62. Muse Spark: 1246
63. Deca 2.6 Pro: 1245
64. MiniMax M3: 1241
65. GLM 5: 1240
66. Gemini 3 Pro Preview: 1238
67. Claude Opus 4.5: 1236
68. GPT-5.6 Terra: 1235
69. DeepSeek-V4-Pro: 1234
70. Beam 1.0 (max): 1233
71. Mercury 3: 1232
72. Qwen3.6 Plus: 1230
73. Kimi K2.5 (Thinking): 1230
74. MiniMax M2.7: 1227
75. Nex N2 Pro: 1226
76. Gemini 3.1 Pro Preview: 1225
77. GPT-5.6 Luna: 1225
78. DeepSeek-V4-Flash-0731: 1224
79. GLM 5V Turbo: 1222
80. Hy3: 1221
81. Grok 4.20 Beta (Reasoning): 1215
82. GLM 4.7: 1211
83. GPT-5.4 (Design Skill, Medium): 1206
84. GPT-5.4 (Medium): 1204
85. Inkling: 1201
86. MiniMax M2.5: 1201
87. Grok 4.20 Beta: 1200
88. DeepSeek-V4-Flash: 1200
89. Deca 2.6 Mini: 1189
90. MiniMax M2.1: 1185
91. Gemini 3 Flash Preview: 1184
92. Grok 4.3: 1181
93. Claude Sonnet 4.5: 1177
94. Step 3.7 Flash: 1176
95. Claude Sonnet 4.5 (Thinking): 1176
96. Solar Pro 4: 1176
97. GPT-5.4 (Low): 1174
98. Qwen3.5 397B A17B: 1174
99. GLM 4.7 Flash: 1172
100. GPT-5.4 (None): 1172
101. Claude 3.7 Sonnet: 1171
102. Hy3 Low: 1171
103. DeepSeek-V3.1 (Thinking): 1169
104. DeepSeek-V3.2-Exp: 1165
105. GPT-5.1 (High): 1165
106. Claude Opus 4.1 (Thinking): 1165
107. GPT-5.2 (Medium): 1164
108. GPT-5.2 (None): 1163
109. GPT-5 (High): 1162
110. Qwen3.5 Plus 02-15: 1161
111. GLM 4.6: 1160
112. GPT-5.2 (Low): 1160
113. GLM 4.5: 1159
114. GPT-5 (Minimal): 1159
115. Claude Opus 4.1: 1159
116. DeepSeek-V3.2: 1158
117. Muse Glimmer 30B (xhigh): 1157
118. GPT-5.1 (Medium): 1155
119. Claude Opus 4: 1155
120. MiMo-V2-Flash: 1150
121. GPT-5.1 (Low): 1150
122. Nemotron 3.5 Super VL: 1150
123. Gemini 2.5 Pro: 1147
124. GPT-5.1 Codex: 1147
125. GPT-5.1 (None): 1145
126. GPT-5.2 (High): 1145
127. Flint: 1143
128. Nemotron 3 Ultra: 1142
129. GPT-5.3 Codex: 1139
130. v0-1.5-lg: 1138
131. Qwen3 Coder 480B A35B Instruct: 1137
132. Mistral Large 3 (2512): 1136
133. Claude Sonnet 4: 1136
134. Solar Mini 4: 1134
135. DeepSeek-R1-0528: 1133
136. Claude Sonnet 4 (Thinking): 1131
137. AGI-01 Swift: 1131
138. GLM 4.5 Air: 1131
139. MiniMax M2 Stable: 1129
140. AesCoder-4B: 1119
141. Mistral Medium 3.5: 1116
142. Mistral Medium 3.1 (2508): 1114
143. Trinity Large Thinking: 1111
144. GPT-5 mini (Default): 1109
145. Claude Haiku 4.5: 1109
146. DeepSeek-V3.1: 1106
147. Qwen3 Max: 1104
148. DeepSeek-V3-0324: 1102
149. Prime Intellect: INTELLECT-3: 1101
150. Gemini 2.5 Flash Preview 09-2025: 1098
151. Grok 4 Fast: 1095
152. Kimi K2 0905 Preview: 1092
153. GPT-5.1 Codex Mini: 1088
154. Grok 4.1 Fast: 1086
155. Grok 4.1 Fast (Reasoning): 1081
156. GPT-5 nano (Default): 1078
157. Kimi K2 Turbo Preview: 1078
158. Gemini 2.5 Flash Lite Preview 09-2025: 1075
159. Gemini 3.1 Flash-Lite Preview: 1066
160. Mistral Medium 3 (2505): 1063
161. Ministral 3 14B (2512): 1059
162. Gemini 2.5 Flash: 1053
163. v0-1.5-md: 1051
164. Grok 3: 1047
165. Ministral 3 8B (2512): 1047
166. Grok 4 Fast (Reasoning): 1044
167. Qwen3-235B-A22B-Instruct-2507: 1033
168. Kimi K2: 1028
169. Qwen3-235B-A22B-Thinking-2507: 1027
170. Magistral Medium 1.2 (2509): 1027
171. GPT-4.1: 1020
172. Mercury 2: 1015
173. o3: 1014
174. Grok 4: 1010
175. Devstral Medium: 1007
176. Ministral 3 3B (2512): 1004
177. Codestral 2508: 1001
178. Qwen3-235B-A22B: 996
179. Grok Code Fast 1: 993
180. GPT-4.1 mini: 988
181. Magistral Small 1.2 (2509): 980
182. o4-mini: 970
183. Olmo 3.1 32B Think: 969
184. GPT OSS 120B: 957
185. GPT-4.1 nano: 957
186. Qwen3 30B-A3B: 936
187. Grok 3 Mini: 925
188. Llama 3.1 Nemotron Ultra 253B: 923
189. Mistral Small 3.2: 901
190. Llama 4 Maverick: 874
191. Mistral Large 2.1 (2411): 857
192. GPT-4o: 855
193. Codestral 2 (2501): 828
194. Devstral Small 1.1: 801
195. Llama 4 Scout: 784

### Design Arena: agentic web app, full-stack ⭐⭐

Blind human votes on agent-built full-stack apps; independent, but a smaller and design-led vote.

Two agents build the same full-stack app (React, Supabase, deployed to Vercel) in a sandbox with 23 tools; people vote for the better one. (updated 2026-11-11)

1. Claude Opus 5.5: 1408
2. GPT-6 Astra (xhigh): 1342
3. Kimi K3.1 Pro: 1330
4. Claude Fable 5.1: 1306
5. Kimi K3: 1296
6. Muse Spark 1.3 Max: 1294
7. Claude Opus 5: 1285
8. Qwen3.8 Max: 1285
9. Muse Spark 1.3 (xhigh): 1275
10. Claude Fable 5: 1261
11. Grok 4.6: 1244
12. GPT-5.6 Sol (xhigh): 1241
13. Claude Opus 4.8: 1230
14. Claude Sonnet 5: 1229
15. Grok 4.5: 1228
16. GLM 5.2: 1221
17. Gemini 3.8 Flash: 1221
18. GPT-6 Sol: 1218
19. Claude Opus 4.6: 1200
20. Claude Sonnet 4.6: 1193
21. Muse Spark 1.2: 1191
22. Gemini 3.7 Flash: 1188
23. MiniMax M3: 1181
24. Muse Spark 1.1: 1178
25. Gemini 3.5 Flash: 1178
26. Qwen3.7 Max: 1174
27. Gemini 3.6 Flash: 1162
28. Kimi K2.7 Code: 1156
29. GLM 5.1: 1148
30. GPT-6 Luna: 1143
31. GPT-5.6 Sol (Medium): 1139
32. Claude Opus 4.5: 1137
33. Kimi K2.6: 1136
34. GLM 5V Turbo: 1130
35. GLM 5: 1103
36. Kimi K2.5 (Thinking): 1094
37. GPT-5.5: 1070
38. Gemini 3 Pro Preview: 1055
39. Gemini 3.1 Pro Preview: 1048
40. Gemini 3 Flash Preview: 1044
41. GPT-5.6 Terra: 1043
42. Grok 4.20 Reasoning: 1035
43. Claude Sonnet 4.5: 1031
44. GLM 4.7: 1030
45. GPT 5.2: 1023
46. GLM 4.6: 1010
47. GPT-5.4: 998
48. Grok Code Fast 1: 990
49. Grok 4.3: 988
50. GPT-5.2 Codex: 973
51. GPT-5.3 Codex: 964

### Design Arena: agentic web app, frontend ⭐⭐

Blind human votes on agent-built front ends; independent, but a smaller and design-led vote.

Same arena for multi-file React front-ends (13 tools). (updated 2026-11-11)

1. GPT-6 Astra (max): 1343
2. Claude Opus 5.5: 1319
3. Claude Fable 5.1: 1318
4. GPT-6 Astra (xhigh): 1315
5. Kimi K3.1 Pro: 1298
6. Inkling 2 Small: 1288
7. Qwen3.8 Max: 1286
8. Muse Spark 1.3 Max: 1284
9. Muse Spark 1.3 (xhigh): 1282
10. Kimi K3: 1282
11. Claude Sonnet 5.5: 1281
12. MiMo-V2.6-Pro: 1276
13. GLM-5.3: 1272
14. GPT-6.1 Sol: 1266
15. Claude Opus 5: 1255
16. Claude Fable 5: 1245
17. Grok 4.6: 1233
18. Ling 3.2 Flash: 1233
19. GPT-5.6 Sol (xhigh): 1227
20. Gemini 3.8 Flash: 1225
21. Claude Sonnet 5: 1224
22. Claude Opus 4.8: 1218
23. GPT-6 Sol: 1216
24. Gemini 3.7 Flash: 1213
25. GLM 5.2: 1207
26. Muse Spark 1.2: 1206
27. Grok 4.5: 1202
28. MiniMax M3: 1194
29. Qwen3.7 Max: 1191
30. Muse Spark 1.1: 1190
31. Claude Opus 4.6: 1189
32. Gemini 3.5 Flash: 1185
33. Gemini 3.6 Flash: 1182
34. Inkling: 1179
35. GPT-5.6 Sol (Medium): 1175
36. Claude Sonnet 4.6: 1174
37. GPT-6 Luna: 1161
38. Kimi K2.7 Code: 1160
39. GLM 5.1: 1160
40. Claude Opus 4.5: 1143
41. Grok 4.20 Reasoning: 1128
42. Kimi K2.5 (Thinking): 1120
43. Gemini 3.1 Pro Preview: 1110
44. Grok 4.3: 1110
45. GLM 5V Turbo: 1109
46. GPT-5.6 Terra: 1103
47. GPT-5.5: 1101
48. Gemini 3 Pro Preview: 1098
49. Gemini 3 Flash Preview: 1094
50. Solar Pro 4: 1089
51. GPT 5.2: 1070
52. Claude Sonnet 4.5: 1044
53. GPT-5.4: 1043
54. GPT-5.2 Codex: 1035
55. GPT-5.3 Codex: 1016
56. GPT-5.1 Codex: 1002

### Design Arena: website ⭐⭐

Blind human votes on websites; independent, but narrow and judged on looks.

Single-shot websites generated by each model, judged by blind votes (1,190,870 votes, Design Arena's largest web board). No standard error is published, so close Elo scores cannot be told apart. (updated 2026-11-11)

1. Muse Wonder (max): 1367
2. Muse Spark 1.3 (xhigh): 1352
3. Muse Spark 1.3 Max: 1343
4. Inkling 2 Small: 1333
5. GPT-6 Astra (max): 1332
6. GPT-6 Astra (xhigh): 1331
7. Kimi K3: 1331
8. Muse Ember: 1330
9. Claude Opus 5.5: 1324
10. Kimi K3.1 Pro: 1320
11. GPT-5.6 Sol (xhigh): 1317
12. Gemini 4 Argon: 1314
13. Claude Sonnet 5.5: 1314
14. Muse Spark 1.2: 1310
15. GPT-6.1 Sol: 1309
16. DeepSeek-V4.1-Flash: 1305
17. Claude Fable 5.1: 1305
18. Claude Opus 5: 1305
19. GPT-5.6 Sol (Medium): 1304
20. MiMo-V2.6-Pro: 1304
21. Claude Fable 5.2: 1303
22. Muse Spark Open (xhigh): 1302
23. Gemini 3.7 Flash: 1300
24. Gemini 3.8 Flash: 1296
25. DeepSeek-V4.1-Flash (1017): 1295
26. Gemini 3.6 Flash: 1294
27. Claude Fable 5: 1291
28. GPT-6.1 Terra: 1288
29. Claude Opus 4.6: 1287
30. GPT-6 Sol: 1285
31. GLM-5.3: 1285
32. Ling-3.1-flash: 1285
33. Grok 4.6: 1283
34. Claude Opus 4.6 (Thinking): 1282
35. Claude Sonnet 4.6: 1281
36. GLM 5.2: 1280
37. Grok 4.5: 1276
38. MiniMax M3.1 Flash: 1275
39. Deca 2.5 Mini: 1274
40. Deca 2.6 Ultra: 1273
41. Qwen3.7 Max: 1271
42. Qwen3.8 Max: 1271
43. Claude Sonnet 5: 1269
44. GLM-5.3-Flash: 1266
45. Muse Spark 1.1: 1266
46. Kimi K2.6: 1265
47. GPT-6 Luna: 1265
48. Gemini 3.5 Flash: 1265
49. Deca 2.5 Ultra: 1265
50. GPT-5.6 Terra: 1264
51. MiMo-V2.5-Pro: 1264
52. Qwen3.7 Plus: 1264
53. MiMo-V2.5: 1263
54. GLM 5.1: 1262
55. Kimi K2.7 Code: 1260
56. GLM 5 Turbo: 1257
57. Claude Haiku 5.5: 1257
58. Muse Spark: 1257
59. Hy4: 1256
60. Claude Opus 4.8: 1255
61. MiniMax M3: 1253
62. Mistral Large 4: 1253
63. GPT-5.5: 1252
64. GPT-5.6 Luna: 1251
65. Gemini 3 Pro Preview: 1249
66. Grok 4.7: 1245
67. Beam 1.0 (max): 1245
68. Kimi K2.5 (Thinking): 1244
69. Claude Opus 4.5: 1243
70. GLM 5: 1243
71. Qwen3.6 Plus: 1238
72. MiniMax M2.7: 1238
73. Mercury 3: 1235
74. Deca 2.6 Pro: 1234
75. DeepSeek-V4-Flash-0731: 1233
76. DeepSeek-V4-Pro: 1230
77. Gemini 3.1 Pro Preview: 1230
78. Hy3: 1226
79. Grok 4.20 Beta (Reasoning): 1225
80. GLM 5V Turbo: 1225
81. GLM 4.7: 1222
82. Inkling: 1221
83. MiniMax M2.5: 1218
84. Nex N2 Pro: 1217
85. Grok 4.20 Beta: 1216
86. GPT-5.4 (Design Skill, Medium): 1216
87. GPT-5.4 (Medium): 1211
88. Deca 2.6 Mini: 1208
89. DeepSeek-V4-Flash: 1203
90. MiniMax M2.1: 1197
91. GPT-5.4 (None): 1195
92. GPT-5.4 (Low): 1194
93. Step 3.7 Flash: 1193
94. Grok 4.3: 1192
95. Gemini 3 Flash Preview: 1192
96. GPT-5.2 (xhigh): 1191
97. GLM 4.7 Flash: 1190
98. Claude 3.7 Sonnet: 1189
99. Qwen3.5 397B A17B: 1187
100. Claude Sonnet 4.5 (Thinking): 1186
101. Claude Sonnet 4.5: 1185
102. DeepSeek-V3.1 (Thinking): 1183
103. Qwen3.5 Plus 02-15: 1183
104. GPT-5.1 (High): 1183
105. GPT-5.2 (None): 1180
106. GPT-5.2 (Medium): 1180
107. GPT-5 (Minimal): 1180
108. Muse Glimmer 30B (xhigh): 1178
109. GPT-5 (High): 1178
110. Hy3 Low: 1178
111. DeepSeek-V3.2-Exp: 1175
112. GPT-5.1 (Medium): 1174
113. Claude Opus 4.1 (Thinking): 1173
114. GPT-5.2 (Low): 1173
115. Solar Pro 4: 1172
116. DeepSeek-V3.2: 1171
117. GLM 4.6: 1170
118. GPT-5.1 (Low): 1168
119. Nemotron 3.5 Super VL: 1166
120. GLM 4.5: 1166
121. GPT-5.1 (None): 1163
122. Gemini 2.5 Pro: 1163
123. MiMo-V2-Flash: 1163
124. Claude Opus 4: 1161
125. Claude Opus 4.1: 1160
126. GPT-5.3 Codex: 1159
127. Mistral Large 3 (2512): 1158
128. Flint: 1158
129. GPT-5.1 Codex: 1158
130. Qwen3 Coder 480B A35B Instruct: 1155
131. GPT-5.2 (High): 1151
132. Solar Mini 4: 1147
133. DeepSeek-R1-0528: 1145
134. v0-1.5-lg: 1142
135. GLM 4.5 Air: 1142
136. Nemotron 3 Ultra: 1142
137. Claude Sonnet 4: 1142
138. MiniMax M2 Stable: 1139
139. Mistral Medium 3.5: 1138
140. AesCoder-4B: 1137
141. Claude Sonnet 4 (Thinking): 1135
142. Trinity Large Thinking: 1133
143. Mistral Medium 3.1 (2508): 1129
144. GPT-5 mini (Default): 1121
145. DeepSeek-V3.1: 1118
146. Claude Haiku 4.5: 1118
147. DeepSeek-V3-0324: 1116
148. Qwen3 Max: 1115
149. Prime Intellect: INTELLECT-3: 1113
150. Gemini 2.5 Flash Preview 09-2025: 1111
151. Grok 4 Fast: 1110
152. Kimi K2 Thinking: 1109
153. GPT-5.1 Codex Mini: 1107
154. AGI-01 Swift: 1105
155. Kimi K2 0905 Preview: 1103
156. Grok 4.1 Fast: 1100
157. Grok 4.1 Fast (Reasoning): 1099
158. GPT-5 nano (Default): 1097
159. Gemini 2.5 Flash Lite Preview 09-2025: 1095
160. Kimi K2 Turbo Preview: 1085
161. Kimi K2 Instruct 0905: 1085
162. Qwen3 Coder 30B A3B Instruct: 1083
163. Gemini 3.1 Flash-Lite Preview: 1079
164. Ministral 3 14B (2512): 1077
165. Mistral Medium 3 (2505): 1075
166. Gemini 2.5 Flash: 1070
167. Ministral 3 8B (2512): 1061
168. v0-1.5-md: 1056
169. Grok 3: 1055
170. Qwen3-235B-A22B-Instruct-2507: 1054
171. Grok 4 Fast (Reasoning): 1051
172. Qwen3-235B-A22B-Thinking-2507: 1049
173. Kimi K2: 1046
174. Magistral Medium 1.2 (2509): 1036
175. GPT-4.1: 1035
176. Mercury 2: 1034
177. o3: 1032
178. Qwen3-235B-A22B: 1026
179. Ministral 3 3B (2512): 1024
180. Devstral Medium: 1015
181. Grok 4: 1010
182. Codestral 2508: 1009
183. GPT-4.1 mini: 993
184. Grok Code Fast 1: 992
185. Magistral Small 1.2 (2509): 988
186. o4-mini: 981
187. Olmo 3.1 32B Think: 974
188. GPT-4.1 nano: 969
189. GPT OSS 120B: 964
190. Qwen3 30B-A3B: 951
191. Llama 3.1 Nemotron Ultra 253B: 931
192. Qwen3 30B-A3B Thinking 2507: 926
193. Grok 3 Mini: 925
194. Magistral Small (2507): 908
195. Mistral Small 3.2: 892
196. Llama 4 Maverick: 867
197. Mistral Large 2.1 (2411): 854
198. Grok 2: 853
199. GPT OSS 20B: 848
200. Amazon Nova Premier: 831
201. GPT-4o: 827
202. Codestral 2 (2501): 805
203. Devstral Small 1.1: 798
204. Amazon Nova Pro: 792
205. Gemini 1.5 Pro: 780
206. Llama 4 Scout: 746

### Arena Code: Image-to-WebDev ⭐⭐

Arena's blind votes, but narrow: turning a screenshot into a web page.

Blind pairwise votes on web pages built from an image, such as a mockup or a screenshot (210,832 votes, cutoff 2026-11-10). (updated 2026-11-10)

1. Claude Opus 5.5 (Max): 1751.9386 ±19.18
2. Claude Sonnet 5.5 (xHigh): 1744.1073 ±14.02
3. Claude Fable 5.1 (Max): 1723.5578 ±11.68
4. GPT-6 Astra (Max): 1722.1349 ±11.26
5. GPT-6.1 Sol (Max): 1695.8773 ±12.91
6. Claude Opus 5 (Max): 1663.9108 ±9.52
7. GPT-6 Sol (Max): 1655.4762 ±10.74
8. Qwen3.8-Max (0902): 1638.1673 ±18.99
9. Muse Spark 1.3 (Max): 1637.3836 ±8.63
10. Claude Fable 5 (High): 1620.5345 ±8.65
11. Qwen3.8-Max: 1615.5279 ±10.18
12. DeepSeek-V4.1-Flash-20260910 (Max): 1610.3578 ±22.57
13. DeepSeek-V4.1-Flash-20261017 (Max): 1609.1731 ±10.45
14. Grok 4.6 (High): 1607.0519 ±10.19
15. GPT-5.6 Sol (xHigh, Codex Harness): 1606.3611 ±9.06
16. Gemini 3.7 Flash (High): 1600.2617 ±11.59
17. Muse Spark 1.3 (xHigh): 1598.2047 ±11.12
18. MiMo-V2.6-Pro: 1596.7082 ±10.77
19. Gemini 3.8 Flash (High): 1593.7005 ±11.25
20. MiMo-V2.6-Flash: 1593.1826 ±11.03
21. Grok 4.7 (xHigh): 1588.7556 ±13.98
22. Kimi K3 (Max): 1586.0086 ±9.6
23. GLM-5.3-Flash: 1582.2689 ±8.08
24. Claude Opus 4.7 (High): 1579.4766 ±9.28
25. Qwen3.8-27B: 1574.3882 ±7.83
26. GPT-6 Luna (Max): 1565.9058 ±8.18
27. Claude Opus 4.7: 1562.1901 ±9
28. Grok 4.5: 1561.6018 ±9.48
29. Step 5 Preview (High): 1551.3546 ±10.35
30. Gemini 3.6 Flash (High): 1544.9113 ±14.56
31. Claude Opus 4.6 (High): 1543.9422 ±9
32. GPT-5.6 Terra (xHigh, Codex Harness): 1539.1346 ±9.08
33. Muse Spark 1.1: 1539.0476 ±11.5
34. Claude Sonnet 5 (High): 1538.8179 ±7.53
35. Claude Opus 4.6: 1538.0466 ±9.11
36. GPT-5.5 (xHigh, Codex Harness): 1530.9548 ±9.27
37. Claude Sonnet 4.6: 1529.5776 ±8.94
38. Kimi K2.6: 1522.1569 ±11.55
39. GPT-5.5 (High, Codex Harness): 1519.441 ±8.97
40. GPT-5.6 Luna (xHigh, Codex Harness): 1517.3625 ±9.94
41. Kimi K2.7 Code: 1512.2969 ±20.42
42. Gemini 3.5 Flash (Medium): 1505.5859 ±11.92
43. Seed2.1 Pro Preview: 1504.8108 ±9.4
44. gemini-3.5-flash: 1504.0808 ±27.07
45. GPT-5.5 (Codex Harness): 1503.5753 ±9.04
46. Gemini 3.5 Flash (High): 1500.2087 ±14.02
47. Gemini 3.1 Pro Preview: 1480.1285 ±8.82
48. MiniMax-M3: 1473.4613 ±8.39
49. Qwen3.6-Plus: 1472.8914 ±10.48
50. Gemini 3 Flash: 1450.9482 ±9.37
51. Gemini 3.5 Flash-Lite: 1449.4359 ±15.59
52. GPT-5.3-Codex (Codex Harness): 1443.7947 ±14.21
53. GPT-5.4: 1441.2416 ±11.55
54. Gemini 3 Pro: 1440.9639 ±22.42
55. Kimi K2.5 (Thinking): 1438.6044 ±17.43
56. Gemini 3 Flash (Minimal): 1431.0489 ±9.3
57. GLM-5V-Turbo: 1427.1935 ±16.2
58. GPT-5.1 (High): 1419.2553 ±21.81
59. Kimi K2.5 (Instant): 1411.2511 ±22.2
60. Inkling-Small: 1409.6306 ±19.52
61. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B (NVFP4): 1391.7801 ±15.15
62. Grok 4.3: 1375.5027 ±9.78
63. Mistral Large 3: 1362.2807 ±17.68
64. GPT-5.1: 1349.0098 ±21.46
65. Gemini 3.1 Flash-Lite Preview: 1338.1526 ±10.97
66. Gemini 2.5 Pro: 1285.2153 ±21.28

### WeirdML v3 (agentic) ⭐⭐

Independent and repeated runs, but machine-learning tasks rather than everyday software work.

Unusual machine-learning problems solved end to end by an agent (Claude Code, Codex CLI, Gemini CLI, Kimi Code, OpenCode), with the cost of each run. Data from Epoch AI. (updated 2026-11-11)

1. Codex - gpt-6-astra (xhigh): 42.2202
2. Codex - gpt-6.1-sol (xhigh): 38.2536
3. Claude Code - claude-opus-5-5 (xhigh): 31.2003
4. Codex - gpt-6.1-terra (xhigh): 28.0863
5. Claude Code - claude-fable-5-1 (xhigh): 25.9702
6. Gemini CLI - gemini-4-argon (high): 20.2617
7. Claude Code - claude-sonnet-5-5 (xhigh): 19.9403
8. Codex - gpt-6-sol (xhigh): 19.7318
9. Claude Code - claude-opus-5 (xhigh): 19.0112
10. Claude Code - claude-fable-5 (xhigh): 15.5869
11. Codex - muse-wonder-1.0 (xhigh): 15.1717
12. Codex - gpt-5.6-sol (xhigh): 14.9922
13. Claude Code - claude-haiku-5-5 (xhigh): 11.6952
14. Kimi Code - kimi-k3.1-pro (xhigh): 9.89
15. Codex - deepseek-v4.1-flash-1017 (high): 9.1909
16. Codex - grok-4.7 (xhigh): 8.4527
17. Codex - gpt-6-luna (xhigh): 7.8741
18. Kimi Code - kimi-k3 (high): 7.5444
19. Gemini CLI - gemini-3.8-flash (high): 7.3787
20. Codex - muse-spark-1.3 (xhigh): 6.655
21. Claude Code - claude-opus-4-5-20251101 (high): 6.5149
22. OpenCode - deepseek-v4.1-flash (high): 5.9463
23. Codex - gpt-5.6-luna (xhigh): 5.158
24. Codex - gpt-5.2-2025-12-11 (xhigh): 5.0079
25. OpenCode - mistral-large-4-0 (high): 4.4459
26. OpenCode - Inkling (xhigh): 3.903

### GSO (software optimization) ⭐⭐

Independent, hack-adjusted, but narrow (performance optimization) and updated less often.

Making real repositories faster without breaking them, in OpenHands, scored after removing reward hacks. Only GPT-6 Astra among current models. (updated 2026-09-27)

1. Claude Fable 5.1 (xhigh): 87.25
2. GPT-6 Astra (xhigh): 77.45
3. Claude Fable 5 (xhigh): 76.47
4. GPT-5.6 Sol (xhigh): 70.59
5. Claude Opus 4.8 (xhigh): 47.06
6. Claude Opus 4.7 (high): 42.16
7. Claude Opus 4.6 (high): 37.25
8. GPT 5.5 (xhigh): 37.25
9. Claude Sonnet 5 (xhigh): 36.27
10. Claude Opus 4.6: 30.39
11. GPT 5.4 (xhigh): 30.39
12. GPT 5.2 (high): 26.47
13. Claude Opus 4.5: 24.5
14. GPT 5.4 (high): 22.55
15. Gemini 3.1 Pro: 21.57
16. Gemini 3 Pro: 17.65
17. GPT 5.1 (high): 12.75
18. Claude Sonnet 4.5: 12.74
19. Gemini 3 Flash: 7.84
20. GPT 5 (high): 5.88
21. Claude Opus 4: 4.9
22. Claude Sonnet 4: 4.9
23. Claude Sonnet 3.5 V2: 4.6
24. O3 (high): 3.92
25. Qwen3 Coder: 3.92
26. Claude Sonnet 3.7: 3.8
27. O4 Mini (high): 3.6
28. Kimi K2 Instruct: 1.96
29. O3 Mini (high): 1.3
30. GLM 4.5 Air: 0.98
31. Gemini 2.5 Pro: 0
32. GPT 4o: 0

### Scale SWE Atlas: codebase Q&A ⭐⭐

Independent (Scale), but a narrow skill (answering questions on a codebase) and not refreshed recently.

Answering deep questions about a real codebase by exploring and running it, in the vendor's agent — one third of the Artificial Analysis Coding Agent Index. (updated 2026-09-09)

1. Opus 5 (Claude Code) xHigh: 63.17 ±5.01
2. Fable-5.1 (Claude Code) xHigh: 59.95 ±4.85
3. GPT 6 Astra (Codex) xHigh: 59.14 ±4.88
4. Opus 4.8 (Claude Code) xhigh: 57.26 ±4.93
5. GLM 5.2 (Mini-SWE-Agent): 48.12 ±5.07
6. Gemini 3.8 Flash (Mini-SWE-Agent): 47.04 ±5.08
7. GPT-5.6-Sol (Codex) xHigh: 46 ±5
8. GPT 5.5 (Codex) xHigh: 45.43 ±5.08
9. Muse Spark 1.1 (Mini-SWE-Agent) xHigh: 42.2 ±5.08
10. GPT 5.4 (Codex) xHigh: 40.8 ±5.1
11. Opus 4.7 (Claude Code): 40.32 ±5.06
12. Fable-5 (Claude Code) xHigh: 39 ±5
13. Gpt 5.4 xHigh (Mini-SWE-Agent): 36.3 ±4.9
14. Opus 4.6 (Claude Code): 33.3 ±5
15. GPT 5.3 (Codex) xHigh: 32.6 ±4.9
16. Sonnet 4.6 (Claude Code): 31.2 ±5
17. Opus 4.6 (Mini-SWE-Agent): 30 ±4.9
18. DeepSeek V4 Pro (Mini-SWE-Agent): 27.15 ±4.74
19. Muse Spark: 24.2 ±4.6
20. Glm 5 (Mini-SWE-Agent): 20.5 ±4.5
21. Gemini 3.1 Pro (Mini-SWE-Agent): 13.5 ±3.9
22. Kimi K2.5 (Mini-SWE-Agent): 13.1 ±4.1
23. Minimax M2.5 (Mini-SWE-Agent): 10.3 ±3.5
24. Gemini 3 Flash (Mini-SWE-Agent): 8.2 ±3.3

### Scale SWE Atlas: test writing ⭐⭐

Independent (Scale), but a narrow skill (writing tests) and not refreshed recently.

Writing tests that catch real bugs in an existing codebase, in the vendor's agent. (updated 2026-09-09)

1. Fable-5.1 (Claude Code) xHigh: 67.04 ±5.33
2. Opus 5 (Claude Code) xHigh: 62.22 ±5.58
3. Fable-5 (Claude Code) xHigh: 55.6 ±5.8
4. Gemini 3.8 Flash (Mini-SWE-Agent): 53.7 ±5.86
5. GPT 6 Astra (Codex) xHigh: 50.74 ±5.91
6. Opus 4.8 (Claude Code) xhigh: 49.63 ±5.93
7. GPT-5.6-Sol (Codex) xHigh: 45.9 ±6
8. GPT-5.4 (Codex) xHigh: 44.36 ±6.04
9. GPT-5.5 (Codex) xHigh: 42.59 ±5.96
10. GLM 5.2 (Mini-SWE-Agent): 41.48 ±5.96
11. Muse Spark 1.1 (Mini-SWE-Agent) xHigh: 41.48 ±5.96
12. GPT-5.4 (Mini-SWE) xHigh: 40 ±6
13. GPT 5.3 (Codex) xHigh: 38.98 ±6.12
14. Opus 4.7 (Claude Code): 38.52 ±5.93
15. Opus-4.6 (Claude Code): 36.67 ±6.63
16. Opus-4.6 (Mini-SWE): 36.08 ±6.02
17. Sonnet-4.6 (Claude Code): 31.76 ±6.24
18. Muse Spark: 31.11 ±5.76
19. Gemini-3-Flash (Mini-SWE): 30.3 ±5.8
20. Gemini-3.1-Pro (Mini-SWE): 29.84 ±5.86
21. Glm-5 (Mini-SWE): 28.74 ±5.76
22. Deepseek V4 Pro (Mini-SWE-Agent): 27.05 ±5.59
23. Kimi-K2.5 (Mini-SWE): 25.77 ±5.63
24. Minimax-M2.5 (Mini-SWE): 18.6 ±5.2

### Scale SWE Atlas: refactoring ⭐⭐

Independent (Scale), but a narrow skill (refactoring) and not refreshed recently.

Refactoring real code without breaking its behavior, in the vendor's agent. (updated 2026-09-09)

1. GPT 6 Astra (Codex) xHigh: 59.05 ±6.43
2. Fable-5.1 (Claude Code) xHigh: 56.67 ±6.52
3. Fable-5 (Claude Code) xHigh: 54.76 ±6.76
4. Opus-4.7 (Claude Code): 48.57 ±6.73
5. Opus 4.8 (Claude Code): 46.67 ±6.75
6. GPT-5.5 (Codex) xHigh: 44.79 ±6.76
7. Gemini 3.8 Flash (Mini-SWE-Agent): 44.76 ±6.76
8. GPT 5.4 (Codex) xHigh: 44.29 ±6.76
9. GLM 5.2 (Mini-SWE-Agent): 42.38 ±6.76
10. GPT 5.3 (Codex) xHigh: 42.38 ±6.76
11. Opus-4.6 (Claude Code): 35.58 ±6.82
12. Gemini-3.1-Pro (Gemini CLI): 33.81 ±6.64
13. Sonnet-4.6 (Claude Code): 32.21 ±6.77
14. Glm-5 (Mini-SWE-Agent): 24.24 ±6.27
15. Kimi-K2.5 (Mini-SWE-Agent): 20.95 ±6
16. Minimax-M2.5 (Mini-SWE-Agent): 19.52 ±5.89
17. Gemini-3-Flash (Mini-SWE-Agent): 10 ±4.8

### Snorkel Agentic Coding 2.0 ⭐⭐

Independent agentic coding tasks with costs, but no effort stated and fewer models.

Terminal-style agentic tasks without network access, 5 attempts each, with the cost per trial. Run by Snorkel AI (a data vendor); no effort stated. (updated 2026-10-23)

1. GPT-6 Astra: 47.6
2. Muse Wonder: 43.6
3. GPT-6.1 Sol: 42.8
4. Gemini 4 Argon: 41.8
5. Opus 5.5: 40.4
6. Fable 5.1: 39.6
7. Opus 5: 38.9
8. Grok 4.6: 33.6
9. Gemini Flash 3.8: 32.6
10. Grok 4.7: 31.9
11. Kimi K3: 26.1
12. Muse Spark 1.3: 25
13. GLM 5.3: 24.9
14. DeepSeek V4 Pro: 14.8
15. Qwen 3.8 Max: 14.7
16. Nemotron 3 Ultra 550B: 5.1

### Snorkel SWE-bench CLI ⭐⭐

Independent multi-file repository changes, but no effort stated and fewer models.

Repository changes that span at least two files, solved from the command line, with the cost per trial. Run by Snorkel AI; no effort stated. (updated 2026-10-23)

1. Opus 5.5: 16.6
2. Gemini 4 Argon: 16.3
3. Fable 5.1: 14.5
4. Opus 5: 14
5. GPT-6 Astra: 12.4
6. Muse Wonder: 11.4
7. GPT-6.1 Sol: 9.8
8. Gemini Flash 3.8: 6.3
9. Grok 4.7: 6.1
10. Grok 4.6: 5.7
11. Qwen 3.8 Max: 4.9
12. Kimi K3: 4.1
13. GLM 5.3: 3.5
14. DeepSeek V4 Pro: 2.3
15. Muse Spark 1.3: 2.2
16. Nemotron 3 Ultra 550B: 0.5

### Vals AI: ProgramBench ⭐⭐

Independent, building a program from scratch, but one effort per model and a narrow task.

Rebuilding real programs from scratch, run by Vals AI with one effort per model, with the cost per test. (updated 2026-11-09)

1. Claude Fable 5.2 (high): 19
2. Claude Opus 5.5 (max): 18.5
3. Claude Fable 5.1 (max): 7
4. Claude Sonnet 5.5 (max): 6.5
5. GPT-6 Astra (max): 5.5
6. GPT-6.1 Sol (max): 3
7. Claude Opus 5 (max): 3
8. Gemini 4 Argon (high): 2.5
9. Claude Haiku 5.5 (max): 2.5
10. Muse Spark 1.3 (max): 2.5
11. Claude Fable 5 (max): 2
12. GPT-6 Sol (max): 2
13. Kimi K3: 2
14. GPT-5.6 Sol (max): 1.5
15. GLM-5.3 (max): 1.5
16. Grok 4.6 (high): 1
17. Claude Opus 4.8 (max): 1
18. Gemini 3.8 Flash (high): 1
19. Grok 4.7 (xhigh): 0.5
20. MiMo-V2.6-Flash: 0.5
21. DeepSeek V4.1 Flash (max): 0.5
22. GPT-5.6 Terra (max): 0.5
23. Muse Spark 1.3 (xhigh): 0.5
24. MiMo-V2.6-Pro: 0.5
25. Inkling 2 Small: 0.5
26. GPT-5.5 (xhigh): 0.5
27. GLM-5.2 (max): 0.5
28. GPT-6 Luna (max): 0.5
29. Muse Wonder Full (max): 0.5
30. GPT-5.4 (high): 0.5
31. Muse Spark 1.2 (xhigh): 0.5
32. Claude Sonnet 4.6 (max): 0.5
33. Inkling Small: 0.5
34. HY4: 0
35. DeepSeek V4 Pro 0813 (max): 0
36. Claude Sonnet 5 (max): 0
37. Qwen3.8 Max (max): 0
38. Gemini 3.7 Flash (high): 0
39. Claude Opus 4.7 (max): 0
40. GPT-5.6 Luna (max): 0
41. DeepSeek V4 Flash 0731 (high): 0
42. GPT-5.4 (xhigh): 0
43. GLM-5.3-Flash (max): 0
44. Grok 4.5 (high): 0
45. GPT-5.4-mini (xhigh): 0
46. Gemini 3.6 Flash (high): 0
47. Muse Spark 1.1 (xhigh): 0
48. Gemini 3.5 Flash (high): 0
49. GLM-5.1: 0
50. Gemini 3.1 Pro Preview (high): 0
51. Qwen3.6 Plus: 0
52. Qwen3.8 27B (xhigh): 0
53. Kimi K2.7 Code: 0
54. Kimi K2.6: 0
55. Nemotron 3 Ultra 550B A55B: 0
56. DeepSeek V4 Pro 0813 (max): 0
57. Gemini 3 Flash (high): 0
58. Gemini 3.5 Flash-Lite (high): 0
59. Claude Haiku 4.5: 0
60. Grok 4.3 (high): 0
61. Nemotron Lightning 3p5 30B A3B: 0
62. Laguna Xs.2: 0
63. Inkling: 0
64. Laguna M.1: 0
65. Gemini 3.1 Flash Lite (high): 0
66. MiniMax-M2.7: 0

### Vals AI: SRE-bench ⭐⭐

Independent and contamination-free, but a narrow skill (reverse engineering) and one effort per model.

Reverse-engineering real, contamination-free binaries, run by Vals AI with one effort per model, with the cost per test. (updated 2026-11-05)

1. GPT-6 Astra (max): 56.87
2. GPT-6.1 Sol (max): 50.763
3. Gemini 4 Argon (high): 44.275
4. Claude Opus 5.5 (max): 33.588
5. GPT Daybreak Blue (max): 31.679
6. GPT-5.6 Sol (max): 30.534
7. GPT-6 Sol (max): 30.534
8. Claude Sonnet 5.5 (max): 30.153
9. Kimi K3.1 Pro (max): 27.099
10. Claude Fable 5.1 (max): 22.901
11. Claude Opus 5.5 Range (max): 21.756
12. Claude Opus 5 (max): 12.214
13. Muse Wonder Full (max): 11.832
14. Gemini 3.8 Flash (high): 11.069
15. GPT-5.6 Terra (max): 9.542
16. GLM-5.3 (max): 9.16
17. Claude Haiku 5.5 (max): 5.725
18. Gemini 3.7 Flash (high): 4.58
19. MiMo-V2.6-Flash: 4.198
20. Grok 4.7 (xhigh): 3.817
21. GPT-5.5 (xhigh): 3.817
22. Muse Spark 1.3 (max): 3.817
23. MiMo-V2.6-Pro: 3.053
24. GPT-6 Luna (max): 2.672
25. HY4: 2.29
26. DeepSeek V4 Pro 0813 (max): 1.908
27. Inkling 2 Small: 1.908
28. GPT-5.6 Luna (max): 1.527
29. DeepSeek V4.1 Flash (max): 0.763
30. GLM-5.3-Flash (max): 0.763
31. Grok 4.5 (high): 0.763
32. Claude Sonnet 5 (max): 0.382
33. Step 5 Preview: 0.382
34. GLM-5.2 (max): 0
35. Muse Spark 1.2 (xhigh): 0
36. Gemini 3.1 Pro Preview (high): 0
37. Inkling: 0

### CWE-bench v2 ⭐⭐

Independent security-fix tasks, recent, but narrow (vulnerabilities).

Finding and fixing real vulnerabilities in real codebases (250 audit-and-patch tasks across 185 CWEs: 100 public, 150 held out), each model in its own agent at high reasoning, graded by deterministic verifiers (pass@1 over 4 rollouts), with the mean cost per rollout. Run by Collinear AI. (updated 2026-11-02)

1. Claude Opus 5.5 (Claude Code): 67.1 ±5.1
2. Muse Wonder (opencode): 66.1 ±5
3. Gemini 4 Argon (Antigravity): 65 ±5.1
4. GPT-6 Astra (Codex): 64.2 ±5.4
5. DeepSeek-V4.1-Flash (1017) (opencode): 57.6 ±5.1
6. Grok 4.7 (opencode): 54.7 ±5.2
7. Hy4 Preview (opencode): 47.3 ±5.3
8. Mistral Large 4 (opencode): 45.6 ±5.3
9. Qwen3.8-Max (0902) (opencode): 45.4 ±4.9
10. Kimi K3.1 Pro (opencode): 45.3 ±5.1
11. GLM-5.3 (opencode): 42 ±4.9

### Endor Labs Agent Security League ⭐⭐

Independent, secure and correct code, but narrow (security) and fewer models.

Does the code an agent writes pass both functional and security tests? Share of tasks that are correct and secure, per agent × model (Claude Code, Codex, Cursor, Muse Code). Run by Endor Labs (a security vendor). (updated 2026-10-24)

1. Claude Code - Claude Fable 5.1: 36.9
2. Codex - GPT-6 Astra: 34.6
3. Codex - GPT-6.1 Sol: 34.1
4. Claude Code - Claude Opus 5.0: 32.4
5. Claude Code - Claude Opus 5.5: 30.7
6. Cursor - Gemini 4 Argon: 27.4
7. Muse Code - Muse Wonder: 25.1
8. Codex - GPT-6 Sol: 25.1
9. Cursor - Claude Fable 5.0: 25.1
10. Cursor - GPT-5.5: 22.3
11. Codex - GPT-5.6 Sol: 20.1
12. Codex - GPT-5.5: 20.1
13. Codex - GPT-5.4: 18.4
14. Claude Code - Claude Fable 5.0: 18.4
15. Claude Code - Claude Sonnet 5.0: 17.3
16. Cursor - Claude Opus 4.7: 16.8
17. Cursor - Gemini 3.5 Flash: 15.6
18. Cursor - Claude Opus 4.8: 15.1
19. Claude Code - Claude Opus 4.7: 14.5
20. Cursor - Gemini 3.1 Pro: 12.8
21. Cursor - GLM-5.2: 12.3
22. Claude Code - Claude Opus 4.8: 11.7
23. Cursor - Composer 2.5: 11.2
24. Cursor - GPT-5.3: 10.6
25. Claude Code - Claude Sonnet 4.6: 8.9
26. Cursor - Claude Opus 4.6: 8.4
27. Claude Code - Claude Opus 4.5: 8.4
28. SWE-Agent - Claude Sonnet 4: 7.8
29. Claude Code - Claude Opus 4.6: 7.3
30. Claude Code - Gemini 3 Pro: 7.3
31. Claude Code - Claude Sonnet 4: 6.1
32. Claude Code - Gemini 2.5 Pro: 5
33. Cursor - Gemini 3 Pro: 4.5
34. SWE-Agent - Gemini 2.5 Pro: 4.5

### Terminal-Bench-Science (Artificial Analysis run) ⭐⭐

Independent run with the same harness for every model, but scientific terminal tasks rather than software work.

Artificial Analysis's own run of Terminal-Bench-Science with the same harness for every model, at several efforts (mean of 3 repeats). (updated 2026-11-09)

1. Claude Fable 5.2 (Max, Default Fallback): 73.3333
2. Claude Fable 5.2 (High, Default Fallback): 70
3. Claude Fable 5.2 (Xhigh, Default Fallback): 67.619
4. Gemini 4 Carbon (High): 64.7619
5. GPT-6 Astra (Max): 63.3333
6. Gemini 4 Carbon (Medium): 62.381
7. Claude Opus 5.5 (Xhigh, Default Fallback): 61.9048
8. Gemini 4 Carbon (Low): 61.9048
9. Claude Fable 5.2 (Medium, Default Fallback): 60.4762
10. GPT-6 Astra (High): 59.5238
11. Claude Opus 5.5 (Max, Default Fallback): 59.0476
12. GPT-6.1 Sol (Max): 58.0952
13. GPT-6.1 Terra (Max): 57.619
14. GPT-6 Astra (Xhigh): 54.7619
15. Claude Fable 5.2 (Low, Default Fallback): 53.8095
16. Claude Sonnet 5.5 (Max, Default Fallback): 53.3333
17. GPT-6 Astra (Medium): 53.3333
18. Gemini 4 Argon (High): 52.8571
19. Claude Sonnet 5.5 (Xhigh, Default Fallback): 52.381
20. GPT-6.1 Sol (Xhigh): 50.4762
21. Claude Opus 5.5 (High, Default Fallback): 49.0476
22. GPT-6.1 Sol (High): 46.1905
23. Claude Fable 5.1 (Max, Default Fallback): 43.3333
24. Claude Opus 5.5 (Medium, Default Fallback): 43.3333
25. GPT-6.1 Terra (Xhigh): 41.9048
26. Claude Fable 5.1 (Xhigh, Default Fallback): 40.4762
27. Claude Fable 5.1 (High, Default Fallback): 39.5238
28. GPT-6 Astra (Low): 37.1429
29. GPT-6.1 Sol (Medium): 37.1429
30. Muse Wonder Full (Max): 37.1429
31. Claude Sonnet 5.5 (High, Default Fallback): 34.2857
32. Kimi K3.1 Pro: 34.2857
33. GPT-6.1 Terra (High): 33.3333
34. GPT-6 Sol (Max): 30
35. Claude Fable 5.1 (Medium, Default Fallback): 29.0476
36. Claude Opus 5 (Max): 28.5714
37. Muse Wonder (Max): 25.7143
38. GPT-6.1 Terra (Medium): 24.7619
39. Claude Opus 5.5 (Low, Default Fallback): 24.2857
40. GPT-6.1 Terra (Non-reasoning): 24.2857
41. GPT-5.6 Sol (Max): 22.381
42. Muse Wonder (High): 21.9048
43. Claude Fable 5.1 (Low, Default Fallback): 21.4286
44. Claude Haiku 5.5 (Max): 20
45. Muse Ember (Max): 20
46. Claude Haiku 5.5 (Xhigh): 18.0952
47. GPT-6 Sol (Xhigh): 17.619
48. Muse Ember (Xhigh): 17.619
49. GPT-6 Sol (High): 17.1429
50. Claude Sonnet 5.5 (Medium, Default Fallback): 16.1905
51. GPT-5.6 Terra (Max): 13.8095
52. DeepSeek V4.1 Flash 1017 (Max): 12.8571
53. Tencent Hy4: 12.8571
54. Qwen3.8 Max (0902): 11.9048
55. Grok 4.7 (Xhigh): 11.4286
56. GPT-6.1 Sol (Low): 10.9524
57. Muse Spark 1.3 (Max): 10.9524
58. Claude Sonnet 5.5 (Low, Default Fallback): 10.4762
59. DeepSeek V4.1 Flash 1017 (High): 10.4762
60. Gemini 3.8 Flash (Medium): 10.4762
61. Claude Haiku 5.5 (High): 10
62. Gemini 3.8 Flash (High): 10
63. GLM-5.3 (Max): 9.5238
64. Grok 4.7 (High): 9.5238
65. DeepSeek V4.1 Flash (Max): 9.0476
66. GPT-6 Luna (Max): 8.5714
67. GPT-6 Sol (Medium): 8.5714
68. Gemini 3.8 Flash (Low): 8.0952
69. Kimi K3 (Max): 8.0952
70. GPT-6.1 Terra (Low): 7.619
71. Claude Haiku 5.5 (Medium): 7.1429
72. Grok 4.6 (High): 6.1905
73. DeepSeek V4 Pro 0813 (Max): 5.7143
74. GPT-6 Sol (Non-reasoning): 5.7143
75. Inkling 2 Small: 5.7143
76. MiMo-V2.6-Pro: 5.7143
77. GPT-6 Luna (Xhigh): 5.2381
78. GLM-5.3-Flash: 4.7619
79. Grok 4.7 (Low): 4.7619
80. Mistral Large 4 Preview: 4.7619
81. Ling 3.1 Flash 1019: 4.2857
82. MiMo-V2.6-Flash: 4.2857
83. Mistral Large 4 (High): 4.2857
84. Agnes 2.5 Pro Theta: 3.8095
85. GLM-5.3 (Low): 3.8095
86. Muse Spark 1.3 (Xhigh): 3.8095
87. GPT-5.6 Luna (Max): 3.3333
88. GPT-6 Sol (Low): 3.3333
89. MiniMax-M3.1-Flash: 3.3333
90. Mistral Large 4 (Medium): 3.3333
91. GPT-6 Luna (High): 2.8571
92. Grok 4.6 (Xhigh): 2.8571
93. DeepSeek V4.1 Flash 1017 (Non-reasoning): 2.381
94. Step 5 Preview: 2.381
95. Claude Haiku 5.5 (Low): 1.9048
96. Muse Spark Open (Xhigh): 1.9048
97. GPT-6 Luna (Medium): 1.4286
98. GPT-6 Luna (Low): 0.9524
99. GPT-6 Luna (Non-reasoning): 0.9524
100. Kimi K3 (Low): 0.9524
101. Nanbeige4.3-33B: 0.9524
102. Qwen3.8 27B (Medium): 0.9524
103. Apodex 1.1: 0.4762
104. Beam 1.0: 0.4762
105. MiniMax-M3: 0.4762
106. Qwen3.8 27B (Low): 0.4762
107. Qwen3.8 27B (Xhigh): 0.4762
108. Claude 4.5 Haiku (Reasoning): 0
109. Command A+: 0
110. DeepSeek V4 Pro 0813 (Non-reasoning): 0
111. DeepSeek V4.1 Flash (Non-reasoning): 0
112. Gemini 3.1 Flash-Lite: 0
113. Gemini 3.5 Flash-Lite: 0
114. Gemma 4 31B (Reasoning): 0
115. Granite 4.2 30B: 0
116. Granite 4.2 3B: 0
117. Granite 4.2 8B: 0
118. Inkling (Xhigh): 0
119. Inkling Small: 0
120. K-EXAONE 2.0: 0
121. K2 Horizon 3.7B: 0
122. K2 Horizon 375B A23B: 0
123. K2 Horizon 7B: 0
124. K2 Horizon MoVA 36B A4B: 0
125. Mistral Medium 3.5: 0
126. Mistral Small 4 (Reasoning): 0
127. Muse Glimmer (High): 0
128. NVIDIA Nemotron 3 Nano 4B: 0
129. Nemotron 3 Ultra 550B A55B (Reasoning): 0
130. Nemotron 3.5 Lightning: 0
131. Nemotron 3.5 Super VL: 0
132. North Mini Code: 0
133. Qwen3.8 27B (Non-reasoning): 0
134. Solar Pro 4: 0
135. gpt-oss-120b (High): 0

### Design Arena: UI components ⭐⭐

Blind human votes on UI components; independent, but narrow and judged on looks.

Blind votes on interface components written as code (114,714 votes). No standard error is published, so close Elo scores cannot be told apart. (updated 2026-11-11)

1. GPT-6 Astra (xhigh): 1399
2. Kimi K3.1 Pro: 1393
3. Claude Sonnet 5.5: 1386
4. Muse Wonder (max): 1376
5. Claude Opus 5.5: 1370
6. GPT-6 Astra (max): 1365
7. Ling-3.1-flash: 1349
8. Muse Spark 1.3 (xhigh): 1347
9. Muse Spark 1.3 Max: 1345
10. GPT-6.1 Sol: 1343
11. Kimi K3: 1341
12. Claude Opus 5: 1340
13. MiMo-V2.6-Pro: 1334
14. Qwen3.8 Max: 1332
15. Claude Fable 5.1: 1332
16. Claude Fable 5.2: 1332
17. Gemini 4 Argon: 1331
18. GPT-5.6 Sol (xhigh): 1325
19. GPT-5.6 Sol (Medium): 1318
20. Claude Fable 5: 1312
21. Muse Spark Open (xhigh): 1310
22. Muse Ember: 1308
23. DeepSeek-V4.1-Flash: 1308
24. Gemini 3.8 Flash: 1307
25. GPT-6.1 Terra: 1302
26. GLM-5.3-Flash: 1300
27. Muse Spark 1.2: 1299
28. Gemini 3.7 Flash: 1292
29. Gemini 3.6 Flash: 1292
30. DeepSeek-V4.1-Flash (1017): 1290
31. GLM 5.2: 1289
32. GLM-5.3: 1289
33. GPT-6 Luna: 1289
34. MiniMax M3.1 Flash: 1288
35. Grok 4.6: 1287
36. GPT-6 Sol: 1284
37. Muse Spark 1.1: 1280
38. Beam 1.0 (max): 1278
39. Claude Opus 4.6: 1277
40. Claude Haiku 5.5: 1276
41. Claude Opus 4.6 (Thinking): 1274
42. Gemini 3 Pro Preview: 1273
43. Grok 4.5: 1273
44. Deca 2.5 Mini: 1272
45. Claude Sonnet 5: 1272
46. Grok 4.7: 1271
47. Mistral Large 4: 1268
48. Gemini 3.5 Flash: 1265
49. Qwen3.7 Max: 1265
50. Claude Sonnet 4.6: 1265
51. GLM 5.1: 1262
52. Kimi K2.6: 1260
53. Deca 2.6 Ultra: 1259
54. GLM 5 Turbo: 1257
55. Qwen3.7 Plus: 1257
56. GPT-5.5: 1254
57. Claude Opus 4.8: 1253
58. GPT-5.6 Terra: 1252
59. Kimi K2.7 Code: 1252
60. MiMo-V2.5: 1252
61. MiMo-V2.5-Pro: 1245
62. GPT-5.6 Luna: 1243
63. MiniMax M3: 1239
64. Muse Spark: 1238
65. Mercury 3: 1235
66. GPT-5.4 (Medium): 1233
67. Deca 2.6 Pro: 1233
68. Claude Opus 4.5: 1232
69. GPT-5.1 Codex: 1231
70. Kimi K2.5 (Thinking): 1230
71. DeepSeek-V4-Flash-0731: 1228
72. Qwen3.6 Plus: 1225
73. GLM 5: 1224
74. Gemini 3.1 Pro Preview: 1220
75. MiniMax M2.1: 1217
76. DeepSeek-V4-Pro: 1215
77. Nex N2 Pro: 1212
78. GPT-5.4 (Low): 1209
79. Deca 2.5 Ultra: 1207
80. GLM 5V Turbo: 1205
81. MiniMax M2.7: 1204
82. Hy3: 1201
83. GLM 4.7 Flash: 1201
84. Muse Glimmer 30B (xhigh): 1199
85. Inkling: 1195
86. GPT-5.4 (None): 1193
87. GLM 4.7: 1193
88. Grok 4.20 Beta (Reasoning): 1193
89. Grok 4.3: 1187
90. GPT-5.2 (Medium): 1184
91. Grok 4.20 Beta: 1182
92. Hy4: 1180
93. DeepSeek-V3.1 (Thinking): 1176
94. GPT-5.2 (High): 1175
95. GPT-5 (High): 1173
96. Step 3.7 Flash: 1173
97. Qwen3.5 Plus 02-15: 1171
98. GPT-5.2 (None): 1169
99. Claude Sonnet 4.5 (Thinking): 1166
100. MiniMax M2.5: 1166
101. GPT-5.1 (Medium): 1160
102. DeepSeek-V3.2-Exp: 1160
103. DeepSeek-V4-Flash: 1158
104. v0-1.5-lg: 1158
105. Claude Opus 4.1 (Thinking): 1158
106. Qwen3.5 397B A17B: 1157
107. Hy3 Low: 1157
108. Claude Sonnet 4.5: 1155
109. GPT-5.2 (Low): 1152
110. Nemotron 3.5 Super VL: 1152
111. GPT-5.1 (High): 1151
112. Claude Opus 4.1: 1149
113. Flint: 1149
114. GPT-5.1 (Low): 1148
115. Claude Opus 4: 1147
116. Claude 3.7 Sonnet: 1145
117. AGI-01 Swift: 1143
118. DeepSeek-V3.2: 1143
119. Solar Pro 4: 1143
120. GLM 4.6: 1141
121. MiMo-V2-Flash: 1138
122. GLM 4.5: 1138
123. GPT-5 (Minimal): 1138
124. Nemotron 3 Ultra: 1137
125. GPT-5.1 (None): 1137
126. Gemini 2.5 Pro: 1133
127. GPT-5.3 Codex: 1131
128. MiniMax M2 Stable: 1130
129. Claude Sonnet 4: 1122
130. Claude Sonnet 4 (Thinking): 1119
131. GLM 4.5 Air: 1117
132. Qwen3 Coder 480B A35B Instruct: 1106
133. GPT-5 mini (Default): 1100
134. Solar Mini 4: 1099
135. Mistral Large 3 (2512): 1096
136. DeepSeek-R1-0528: 1096
137. Claude Haiku 4.5: 1093
138. Mistral Medium 3.1 (2508): 1093
139. DeepSeek-V3-0324: 1087
140. Gemini 2.5 Flash Preview 09-2025: 1086
141. AesCoder-4B: 1080
142. DeepSeek-V3.1: 1079
143. GPT-5.1 Codex Mini: 1075
144. Mistral Medium 3.5: 1072
145. Qwen3 Max: 1072
146. Prime Intellect: INTELLECT-3: 1071
147. Grok 4 Fast: 1066
148. Gemini 3.1 Flash-Lite Preview: 1065
149. GPT-5 nano (Default): 1060
150. Grok 4.1 Fast (Reasoning): 1047
151. Qwen3 Coder 30B A3B Instruct: 1040
152. Trinity Large Thinking: 1039
153. Grok 4.1 Fast: 1030
154. Grok 3: 1030
155. Kimi K2: 1024
156. Grok 4 Fast (Reasoning): 1021
157. Gemini 2.5 Flash Lite Preview 09-2025: 1019
158. Mistral Medium 3 (2505): 1018
159. Magistral Medium 1.2 (2509): 1017
160. Codestral 2508: 1011
161. o3: 1010
162. v0-1.5-md: 1003
163. Gemini 2.5 Flash: 999
164. GPT-4.1: 994
165. Mercury 2: 991
166. Grok Code Fast 1: 984
167. Grok 4: 980
168. o4-mini: 977
169. Devstral Medium: 973
170. Qwen3-235B-A22B-Instruct-2507: 959
171. GPT-4.1 mini: 955
172. Qwen3-235B-A22B: 953
173. Qwen3 30B-A3B: 939
174. Qwen3-235B-A22B-Thinking-2507: 939
175. Magistral Small 1.2 (2509): 924
176. GPT OSS 120B: 920
177. GPT-4.1 nano: 910
178. Mistral Small 3.2: 900
179. Llama 4 Maverick: 893
180. Grok 3 Mini: 883
181. GPT-4o: 879
182. Olmo 3.1 32B Think: 852
183. Codestral 2 (2501): 846
184. Devstral Small 1.1: 832
185. Mistral Large 2.1 (2411): 818
186. Llama 4 Scout: 761

### Design Arena: game dev ⭐⭐

Blind human votes on small games; independent, but narrow.

Blind votes on browser games written as code (318,370 votes) — the closest board to 3D and creative coding. No standard error is published, so close Elo scores cannot be told apart. (updated 2026-11-11)

1. GPT-6 Astra (max): 1466
2. GPT-6 Astra (xhigh): 1457
3. Claude Opus 5.5: 1429
4. Inkling 2 Small: 1424
5. GPT-6.1 Sol: 1409
6. Muse Wonder (max): 1407
7. Kimi K3.1 Pro: 1406
8. Claude Fable 5.2: 1404
9. Claude Sonnet 5.5: 1394
10. Claude Fable 5.1: 1368
11. Gemini 4 Argon: 1361
12. GPT-5.6 Sol (xhigh): 1360
13. Kimi K3: 1348
14. Muse Spark 1.3 Max: 1338
15. Claude Opus 5: 1338
16. Claude Fable 5: 1329
17. Muse Ember: 1329
18. GPT-6.1 Terra: 1328
19. Muse Spark 1.3 (xhigh): 1325
20. DeepSeek-V4.1-Flash: 1324
21. GPT-5.6 Sol (Medium): 1317
22. Claude Haiku 5.5: 1315
23. Gemini 3.7 Flash: 1313
24. MiMo-V2.6-Pro: 1310
25. Grok 4.7: 1308
26. MiniMax M3.1 Flash: 1308
27. Qwen3.8 Max: 1308
28. Ling-3.1-flash: 1306
29. GPT-6 Sol: 1304
30. Gemini 3.8 Flash: 1298
31. Muse Spark 1.2: 1294
32. GLM-5.3: 1293
33. Grok 4.6: 1289
34. GPT-5.5: 1284
35. Claude Opus 4.6 (Thinking): 1271
36. Claude Sonnet 5: 1271
37. GLM-5.3-Flash: 1270
38. MiMo-V2.5-Pro: 1266
39. Claude Opus 4.6: 1265
40. Muse Spark 1.1: 1263
41. DeepSeek-V4.1-Flash (1017): 1261
42. GLM 5.2: 1260
43. Qwen3.7 Max: 1259
44. GLM 5.1: 1259
45. Gemini 3.5 Flash: 1258
46. Grok 4.5: 1257
47. Gemini 3.6 Flash: 1257
48. Claude Opus 4.8: 1257
49. Hy4: 1255
50. Qwen3.7 Plus: 1255
51. GPT-6 Luna: 1254
52. Muse Spark Open (xhigh): 1254
53. Deca 2.5 Mini: 1253
54. Claude Sonnet 4.6: 1252
55. Deca 2.5 Ultra: 1250
56. Deca 2.6 Ultra: 1250
57. MiMo-V2.5: 1247
58. Kimi K2.6: 1244
59. GLM 5 Turbo: 1241
60. Deca 2.6 Pro: 1235
61. GPT-5.4 (Medium): 1232
62. GLM 5: 1228
63. GPT-5.6 Terra: 1223
64. DeepSeek-V4-Pro: 1223
65. Claude Opus 4.5: 1223
66. Mistral Large 4: 1219
67. Nex N2 Pro: 1219
68. Muse Spark: 1218
69. Kimi K2.7 Code: 1215
70. Mercury 3: 1214
71. GLM 5V Turbo: 1210
72. Qwen3.6 Plus: 1206
73. Hy3: 1205
74. MiniMax M3: 1203
75. MiniMax M2.7: 1202
76. Beam 1.0 (max): 1202
77. DeepSeek-V4-Flash-0731: 1201
78. Kimi K2.5 (Thinking): 1200
79. Gemini 3.1 Pro Preview: 1196
80. DeepSeek-V4-Flash: 1191
81. Gemini 3 Pro Preview: 1190
82. GPT-5.2 (Low): 1189
83. GPT-5.6 Luna: 1188
84. GPT-5.4 (None): 1188
85. Grok 4.20 Beta (Reasoning): 1186
86. GPT-5.4 (Low): 1185
87. GLM 4.7: 1182
88. GPT-5.2 (High): 1180
89. GPT-5 (High): 1179
90. GPT-5.2 (Medium): 1177
91. GPT-5.1 (High): 1172
92. AGI-01 Swift: 1171
93. MiniMax M2.5: 1170
94. GPT-5.2 (None): 1169
95. Grok 4.20 Beta: 1167
96. Grok 4.3: 1165
97. Claude Opus 4.1 (Thinking): 1165
98. Claude Opus 4: 1164
99. Solar Pro 4: 1163
100. Inkling: 1161
101. Claude Sonnet 4.5: 1160
102. Gemini 3 Flash Preview: 1160
103. Claude Sonnet 4.5 (Thinking): 1157
104. Claude Opus 4.1: 1157
105. GPT-5.3 Codex: 1155
106. GPT-5.1 (Medium): 1153
107. Step 3.7 Flash: 1151
108. GPT-5 (Minimal): 1148
109. GPT-5.1 (Low): 1145
110. GLM 4.6: 1144
111. GLM 4.5: 1138
112. DeepSeek-V3.2-Exp: 1138
113. Nemotron 3 Ultra: 1134
114. Qwen3.5 397B A17B: 1134
115. Claude 3.7 Sonnet: 1132
116. GPT-5.1 Codex: 1132
117. Claude Sonnet 4 (Thinking): 1132
118. Deca 2.6 Mini: 1130
119. Hy3 Low: 1129
120. GLM 4.7 Flash: 1128
121. MiniMax M2.1: 1128
122. Claude Sonnet 4: 1127
123. GPT-5.1 (None): 1126
124. DeepSeek-V3.2: 1125
125. MiMo-V2-Flash: 1124
126. DeepSeek-V3.1 (Thinking): 1124
127. GPT-5 mini (Default): 1123
128. v0-1.5-lg: 1122
129. Muse Glimmer 30B (xhigh): 1113
130. MiniMax M2 Stable: 1111
131. Solar Mini 4: 1107
132. Gemini 2.5 Pro: 1105
133. Nemotron 3.5 Super VL: 1103
134. Qwen3.5 Plus 02-15: 1099
135. Qwen3 Coder 480B A35B Instruct: 1093
136. DeepSeek-R1-0528: 1091
137. Claude Haiku 4.5: 1091
138. GLM 4.5 Air: 1088
139. Qwen3 Max: 1088
140. GPT-5.1 Codex Mini: 1086
141. Flint: 1080
142. Kimi K2 Turbo Preview: 1079
143. DeepSeek-V3.1: 1078
144. AesCoder-4B: 1075
145. Mistral Large 3 (2512): 1074
146. Mistral Medium 3.5: 1073
147. Trinity Large Thinking: 1073
148. GPT-4.1: 1073
149. Prime Intellect: INTELLECT-3: 1070
150. Mistral Medium 3.1 (2508): 1068
151. GPT-4.1 mini: 1064
152. Grok 4.1 Fast: 1061
153. Grok 4 Fast: 1060
154. Gemini 2.5 Flash Preview 09-2025: 1059
155. Grok 4.1 Fast (Reasoning): 1054
156. DeepSeek-V3-0324: 1047
157. Gemini 2.5 Flash Lite Preview 09-2025: 1041
158. GPT-5 nano (Default): 1038
159. Ministral 3 14B (2512): 1032
160. o3: 1028
161. Gemini 3.1 Flash-Lite Preview: 1023
162. Grok 3: 1020
163. v0-1.5-md: 1017
164. Grok 4 Fast (Reasoning): 1015
165. Mistral Medium 3 (2505): 1011
166. Gemini 2.5 Flash: 998
167. o4-mini: 997
168. Grok 4: 995
169. Magistral Medium 1.2 (2509): 989
170. GPT OSS 120B: 986
171. Ministral 3 8B (2512): 984
172. Grok Code Fast 1: 982
173. Devstral Medium: 979
174. Mercury 2: 974
175. Kimi K2: 968
176. GPT-4.1 nano: 964
177. Codestral 2508: 961
178. Olmo 3.1 32B Think: 960
179. Qwen3-235B-A22B-Thinking-2507: 951
180. Qwen3-235B-A22B-Instruct-2507: 945
181. Grok 3 Mini: 943
182. Ministral 3 3B (2512): 942
183. Magistral Small 1.2 (2509): 936
184. Qwen3-235B-A22B: 921
185. GPT-4o: 900
186. Qwen3 30B-A3B: 892
187. Mistral Small 3.2: 882
188. Mistral Large 2.1 (2411): 881
189. Llama 3.1 Nemotron Ultra 253B: 863
190. Llama 4 Maverick: 831
191. Codestral 2 (2501): 805
192. Devstral Small 1.1: 774
193. Llama 4 Scout: 766

### Design Arena: mobile apps (agents) ⭐⭐

Blind human votes on agent-built mobile apps; independent, but narrow and a smaller vote.

Two agents build the same mobile app and people vote for the better one (45,185 votes): it compares agents, not bare models. No standard error is published, so close Elo scores cannot be told apart. (updated 2026-11-11)

1. Claude Fable 5.2: 1338
2. Claude Opus 5.5: 1298
3. GPT-6 Astra (xhigh): 1289
4. Kimi K3.1 Pro: 1273
5. Claude Sonnet 5.5: 1263
6. Claude Fable 5: 1259
7. Kimi K3: 1258
8. Claude Fable 5.1: 1258
9. Grok 4.7: 1252
10. GPT-5.6 Sol (xhigh): 1230
11. Gemini 3.7 Flash: 1229
12. Gemini 3.8 Flash: 1229
13. GPT-6.1 Sol: 1228
14. Grok 4.6: 1227
15. Qwen3.8 Max: 1224
16. MiniMax M3.1 Flash: 1220
17. Muse Spark 1.3 Max: 1214
18. Claude Opus 4.8: 1212
19. Muse Spark 1.3 (xhigh): 1210
20. Grok 4.5: 1209
21. Claude Sonnet 4.6: 1207
22. Claude Sonnet 5: 1206
23. Claude Opus 4.6: 1203
24. GPT-6 Sol: 1197
25. Gemini 3.6 Flash: 1194
26. Ling-3.1-flash: 1193
27. GLM-5.3: 1187
28. Claude Opus 4.5: 1184
29. Muse Spark 1.2: 1178
30. GPT-6.1 Terra: 1178
31. Muse Spark 1.1: 1172
32. Gemini 3.5 Flash: 1172
33. MiniMax M3: 1172
34. GLM 5.2: 1157
35. GLM 5.1: 1156
36. Qwen3.7 Max: 1153
37. Kimi K2.6: 1151
38. Kimi K2.5 (Thinking): 1143
39. GPT-5.5: 1138
40. Kimi K2.7 Code: 1138
41. Claude Sonnet 4.5: 1134
42. GLM 5: 1132
43. GLM 5V Turbo: 1131
44. GPT-6 Luna: 1121
45. GPT-5.6 Terra: 1101
46. GLM 4.7: 1101
47. GPT-5.6 Sol (Medium): 1100
48. GLM 4.6: 1092
49. Gemini 3.1 Pro Preview: 1090
50. GPT 5.2: 1089
51. Gemini 3 Flash Preview: 1086
52. Grok 4.20 Reasoning: 1085
53. GPT-5.2 Codex: 1085
54. GPT-5.4: 1082
55. Gemini 3 Pro Preview: 1071
56. Grok 4.3: 1053
57. GPT-5.3 Codex: 1049
58. Grok Code Fast 1: 1029

### Design Arena: SVG ⭐⭐

Blind human votes on SVG drawings; independent, but a very narrow skill.

Blind votes on SVG icons, illustrations and logos written as code (211,612 votes). Includes SVG-specialized models that no subscription offers. (updated 2026-11-11)

1. Arrow 2 Telos: 1500
2. GPT-6 Astra (xhigh): 1421
3. GPT-6 Astra (max): 1411
4. Arrow 2: 1373
5. Gemini 4 Argon: 1329
6. Kimi K3.1 Pro: 1325
7. Claude Opus 5.5: 1320
8. Muse Spark 1.3 Max: 1319
9. Claude Opus 5: 1306
10. GPT-5.6 Sol (xhigh): 1305
11. MiMo-V2.6-Pro: 1303
12. GPT-6 Sol: 1302
13. Claude Sonnet 5.5: 1302
14. Muse Spark 1.3 (xhigh): 1300
15. Kimi K3: 1300
16. Claude Fable 5: 1296
17. Claude Fable 5.1: 1294
18. GPT-5.6 Sol (Medium): 1294
19. Gemini 3.8 Flash: 1290
20. MiniMax M3.1 Flash: 1285
21. Gemini 3.7 Flash: 1280
22. Gemini 3.1 Pro Preview: 1279
23. Muse Spark 1.2: 1275
24. GLM-5.3-Flash: 1273
25. Grok 4.7: 1272
26. Arrow 1.1 Max: 1270
27. Ling-3.1-flash: 1270
28. Arrow-1.0: 1268
29. GLM-5.3: 1255
30. DeepSeek-V4.1-Flash (1017): 1252
31. Gemini 3.5 Flash: 1249
32. GPT-6 Luna: 1241
33. Gemini 3 Pro Preview: 1239
34. DeepSeek-V4.1-Flash: 1239
35. GPT-5.5: 1230
36. Grok 4.6: 1229
37. Mistral Large 4: 1228
38. Muse Spark 1.1: 1227
39. Claude Opus 4.6 (Thinking): 1222
40. Recraft V4.1 Vector Pro: 1221
41. Muse Spark: 1221
42. Recraft V4 Vector: 1214
43. GPT-5.6 Terra: 1213
44. GLM 5.1: 1212
45. Recraft V4 Vector Pro: 1212
46. Qwen3.7 Max: 1211
47. Claude Opus 4.6: 1210
48. Qwen3.7 Plus: 1207
49. Grok 4.5: 1204
50. GLM 5.2: 1204
51. Arrow 1.1: 1199
52. GLM 5 Turbo: 1199
53. Claude Haiku 5.5: 1196
54. Hy3: 1194
55. Nex N2 Pro: 1191
56. Claude Sonnet 4.6: 1191
57. Claude Sonnet 5: 1191
58. GPT-5.4 (Medium): 1189
59. Claude Opus 4.8: 1189
60. GPT-5.6 Luna: 1188
61. MiMo-V2.5-Pro: 1187
62. GPT-5 (High): 1183
63. MiMo-V2.5: 1180
64. DeepSeek-V4-Flash-0731: 1177
65. Kimi K2.6: 1176
66. Claude Opus 4.5: 1175
67. Mercury 3: 1172
68. Kimi K2.7 Code: 1170
69. GLM 5: 1163
70. Grok 4.20 Beta (Reasoning): 1158
71. Qwen3.6 Plus: 1156
72. MiniMax M3: 1154
73. DeepSeek-V4-Flash: 1150
74. Claude Opus 4.1 (Thinking): 1147
75. GLM 5V Turbo: 1145
76. Kimi K2.5 (Thinking): 1144
77. MiniMax M2.5: 1144
78. GPT-5.1 (High): 1143
79. GLM 4.7: 1137
80. GPT-5.4 (Low): 1136
81. Qwen3.5 397B A17B: 1134
82. DeepSeek-V4-Pro: 1134
83. GPT-5.2 (Medium): 1133
84. GPT-5.2 (Low): 1130
85. MiniMax M2.7: 1130
86. MiniMax M2.1: 1126
87. GPT-5.1 (Medium): 1125
88. GPT-5.2 (None): 1125
89. GPT-5.3 Codex: 1123
90. GPT-5 (Minimal): 1122
91. Claude Opus 4: 1122
92. GPT-5.4 (None): 1121
93. GPT-5.1 (Low): 1112
94. Claude Sonnet 4.5 (Thinking): 1110
95. Inkling: 1108
96. Claude Opus 4.1: 1104
97. GLM 4.6: 1103
98. Claude Sonnet 4.5: 1102
99. Qwen3.5 Plus 02-15: 1102
100. GPT-5.1 (None): 1099
101. MiniMax M2 Stable: 1094
102. GLM 4.5: 1091
103. GPT-5 mini (Default): 1086
104. Claude 3.7 Sonnet: 1082
105. Step 3.7 Flash: 1079
106. Grok 4.3: 1076
107. Claude Sonnet 4 (Thinking): 1073
108. GLM 4.5 Air: 1066
109. Nemotron 3 Ultra: 1064
110. DeepSeek-V3.1 (Thinking): 1059
111. Claude Sonnet 4: 1058
112. v0-1.5-md: 1056
113. Grok 4.20 Beta: 1052
114. Gemini 3.1 Flash-Lite Preview: 1047
115. DeepSeek-R1-0528: 1033
116. GLM 4.7 Flash: 1032
117. DeepSeek-V3.2: 1027
118. DeepSeek-V3.2-Exp: 1026
119. MiMo-V2-Flash: 1024
120. Claude Haiku 4.5: 1020
121. Grok 4: 1015
122. Gemini 2.5 Flash Preview 09-2025: 1014
123. Trinity Large Thinking: 1011
124. Qwen3 Max: 1006
125. Grok 4.1 Fast (Reasoning): 1003
126. Mercury 2: 1000
127. Mistral Large 3 (2512): 987
128. Mistral Medium 3.1 (2508): 987
129. Prime Intellect: INTELLECT-3: 986
130. Grok 3: 977
131. GPT-5.1 Codex Mini: 972
132. DeepSeek-V3-0324: 971
133. Grok 4 Fast (Reasoning): 970
134. DeepSeek-V3.1: 961
135. Mistral Medium 3.5: 944
136. Grok 4.1 Fast: 938
137. Grok 4 Fast: 933
138. AesCoder-4B: 926
139. Olmo 3.1 32B Think: 866

### Design Arena: data visualization ⭐⭐

Blind human votes on charts; independent, but a very narrow skill.

Blind votes on charts and dashboards written as code (90,018 votes; no standard error published). (updated 2026-11-11)

1. Claude Opus 5.5: 1390
2. Claude Sonnet 5.5: 1362
3. Kimi K3: 1348
4. Claude Fable 5.1: 1342
5. Muse Spark 1.3 (xhigh): 1340
6. MiMo-V2.6-Pro: 1336
7. Gemini 4 Argon: 1333
8. Claude Opus 5: 1327
9. Muse Spark 1.2: 1326
10. GPT-6 Astra (max): 1325
11. GPT-6 Astra (xhigh): 1325
12. GPT-5.6 Sol (xhigh): 1320
13. Gemini 3.7 Flash: 1314
14. Muse Ember: 1310
15. Muse Spark 1.3 Max: 1307
16. Kimi K3.1 Pro: 1306
17. Muse Wonder (max): 1305
18. Claude Fable 5: 1302
19. GPT-5.6 Sol (Medium): 1290
20. Gemini 3.6 Flash: 1288
21. DeepSeek-V4.1-Flash: 1288
22. Ling-3.1-flash: 1286
23. Qwen3.8 Max: 1284
24. Claude Haiku 5.5: 1284
25. GLM 5.2: 1283
26. GPT-6.1 Sol: 1281
27. Grok 4.6: 1281
28. Claude Opus 4.6: 1277
29. Claude Sonnet 4.6: 1276
30. Qwen3.7 Plus: 1276
31. Gemini 3.8 Flash: 1273
32. Mistral Large 4: 1272
33. Qwen3.7 Max: 1272
34. GPT-6 Sol: 1271
35. Deca 2.5 Ultra: 1270
36. Claude Opus 4.6 (Thinking): 1267
37. Muse Spark 1.1: 1265
38. Grok 4.5: 1265
39. DeepSeek-V4.1-Flash (1017): 1264
40. GLM 5 Turbo: 1261
41. GPT-5.5: 1260
42. GLM-5.3-Flash: 1259
43. GLM 5.1: 1259
44. Muse Spark: 1258
45. MiMo-V2.5-Pro: 1256
46. GLM-5.3: 1252
47. Claude Sonnet 5: 1250
48. MiMo-V2.5: 1248
49. Kimi K2.6: 1247
50. Claude Opus 4.5: 1241
51. Gemini 3.5 Flash: 1238
52. Gemini 3 Pro Preview: 1237
53. GPT-5.6 Terra: 1237
54. Claude Opus 4.8: 1236
55. GPT-5.4 (Medium): 1233
56. Deca 2.5 Mini: 1233
57. MiniMax M3: 1231
58. MiniMax M2.7: 1230
59. Grok 4.7: 1228
60. GLM 5: 1226
61. Qwen3.6 Plus: 1225
62. Nex N2 Pro: 1223
63. Kimi K2.7 Code: 1221
64. Gemini 2.5 Pro: 1220
65. GPT-5 (High): 1219
66. Kimi K2.5 (Thinking): 1217
67. MiniMax M3.1 Flash: 1215
68. Grok 4.20 Beta (Reasoning): 1215
69. Mercury 3: 1212
70. GPT-5.1 Codex: 1211
71. GPT-6 Luna: 1210
72. GPT-5.6 Luna: 1207
73. Gemini 3.1 Pro Preview: 1207
74. MiniMax M2.1: 1204
75. GPT-5.1 (Low): 1202
76. DeepSeek-V4-Pro: 1200
77. GLM 5V Turbo: 1200
78. Grok 4.20 Beta: 1197
79. Claude 3.7 Sonnet: 1194
80. GPT-5.4 (Low): 1194
81. GLM 4.7: 1193
82. GPT-5.1 (None): 1193
83. GPT-5.1 (High): 1192
84. GPT-5.1 (Medium): 1190
85. GLM 4.5 Air: 1184
86. GPT-5.2 (Low): 1184
87. GPT-5.2 (None): 1182
88. DeepSeek-R1-0528: 1180
89. GPT-5.4 (None): 1179
90. DeepSeek-V4-Flash-0731: 1178
91. Grok 4.3: 1178
92. GPT-6.1 Terra: 1176
93. Solar Pro 4: 1175
94. GPT-5 (Minimal): 1174
95. Qwen3.5 397B A17B: 1173
96. Solar Mini 4: 1173
97. Step 3.7 Flash: 1171
98. MiniMax M2.5: 1170
99. DeepSeek-V3.1 (Thinking): 1166
100. Claude Sonnet 4.5: 1165
101. Hy3: 1164
102. Claude Opus 4.1 (Thinking): 1164
103. GPT-5.3 Codex: 1162
104. Claude Opus 4.1: 1159
105. GLM 4.6: 1159
106. GLM 4.5: 1157
107. DeepSeek-V3.2: 1156
108. Inkling: 1156
109. Claude Sonnet 4.5 (Thinking): 1156
110. Claude Sonnet 4 (Thinking): 1155
111. GPT-5.2 (Medium): 1154
112. DeepSeek-V3.2-Exp: 1150
113. Claude Opus 4: 1144
114. Nemotron 3 Ultra: 1143
115. Mistral Medium 3.1 (2508): 1142
116. MiniMax M2 Stable: 1138
117. MiMo-V2-Flash: 1135
118. Mistral Large 3 (2512): 1134
119. Qwen3.5 Plus 02-15: 1133
120. Gemini 2.5 Flash Preview 09-2025: 1132
121. GPT-5 mini (Default): 1129
122. DeepSeek-V4-Flash: 1125
123. Hy3 Low: 1123
124. Claude Haiku 4.5: 1122
125. GLM 4.7 Flash: 1121
126. Mistral Medium 3.5: 1118
127. Claude Sonnet 4: 1110
128. AesCoder-4B: 1108
129. Nemotron 3.5 Super VL: 1099
130. Qwen3 Max: 1098
131. Trinity Large Thinking: 1096
132. Grok 4 Fast: 1095
133. GPT-5.1 Codex Mini: 1095
134. DeepSeek-V3.1: 1094
135. DeepSeek-V3-0324: 1087
136. Grok 4.1 Fast (Reasoning): 1086
137. AGI-01 Swift: 1082
138. Prime Intellect: INTELLECT-3: 1076
139. Qwen3-235B-A22B-Instruct-2507: 1068
140. Grok 4.1 Fast: 1062
141. Gemini 2.5 Flash: 1058
142. GPT-5 nano (Default): 1054
143. Gemini 3.1 Flash-Lite Preview: 1043
144. Grok 4 Fast (Reasoning): 1037
145. Mistral Medium 3 (2505): 1032
146. Grok 4: 1029
147. Grok 3: 1028
148. Codestral 2508: 1015
149. Mercury 2: 1011
150. o4-mini: 989
151. GPT OSS 120B: 985
152. v0-1.5-md: 983
153. Devstral Medium: 933
154. Olmo 3.1 32B Think: 901
155. GPT-4.1 nano: 889
156. Grok 3 Mini: 875

### llm-stats: Best AI for coding ⭐

An aggregate that mixes many scores, most of them self-reported by the vendors: useful context, not a measurement.

Aggregate rating across 68 coding benchmarks, mostly self-reported by vendors. (updated 2026-11-11)

1. Claude Fable 5.2: 50.7
2. Claude Opus 5.5: 48.8
3. Claude Sonnet 5.5: 47.1
4. Claude Mythos 5.2: 45.1
5. GPT-6 Astra: 45.1
6. Claude Mythos Preview: 43.2
7. Claude Fable 5: 43
8. Muse Wonder Full: 43
9. Gemini 4 Argon: 42.3
10. Muse Wonder Full: 42
11. Kimi K3.1 Pro: 41.7
12. GPT-5.6 Sol: 41.6
13. Muse Wonder: 40.4
14. DeepSeek-V4.1-Flash: 40.1
15. Kimi K3: 40
16. DeepSeek-V4.1-Flash-1017: 39.8
17. Muse Wonder: 39.3
18. GLM-5.3: 39.1
19. Muse Spark 1.3: 39
20. GPT-6.1 Sol: 38.8
21. Claude Opus 5: 38.8
22. Claude Opus 4.8: 38.4
23. Claude Haiku 5.5: 38.4
24. MiMo-V2.6-Pro: 38.3
25. GPT-5.6 Terra: 38.3
26. DeepSeek-V4-Pro-0813: 38.1
27. LongCat-2.5: 38
28. MiniMax M3.1 Flash: 38
29. SWE-2: 38
30. Qwen3.8 Max: 37.2
31. Ember-1: 36.8
32. Hy4 preview: 36.6
33. Mistral Large 4: 36.4
34. Inkling 2 Small: 36.2
35. GPT-6.1 Terra: 35.9
36. GPT-5.5: 35.5
37. Claude Mythos 5.1: 35.2
38. Grok 4.7: 34.7
39. Claude Opus 4.7: 34.7
40. Claude Fable 5.1: 34.6
41. Gemini 3.8 Flash: 34.3
42. Qwen3.7 Max: 34.1
43. Claude Sonnet 5: 33.9
44. MiMo-V2.6-Flash: 33.8
45. Muse Spark Open: 33.5
46. GPT-5.6 Luna: 33.3
47. Beam 1.0: 33.3
48. GPT-6 Sol: 33.3
49. GLM-5.2: 33.1
50. Qwen3.8-Flash-Next: 32.9
51. Qwen3.8 Flash: 32.8
52. DeepSeek-V4-Flash-Vision-Exp: 32.7
53. Kimi K2.6: 32.6
54. Grok 4.6: 32.6
55. Muse Spark 1.1: 32.1
56. Seed 2.1 Pro: 31.9
57. Gemini 3.7 Flash: 31.7
58. Grok 4.5: 31.7
59. GLM-5.3-Flash: 31.5
60. MiMo-V2.5-Pro: 31.2
61. Hy3: 31.1
62. Claude Opus 4.6: 31
63. DeepSeek-V4-Pro-Max: 30.9
64. DeepSeek-V4-Flash-0731: 30.6
65. MiniMax M3: 30.1
66. Gemini 3.1 Pro: 29.8
67. Qwen3.7-Plus: 29.4
68. Gemini 3.5 Flash: 29.3
69. Seed 2.1 Turbo: 29.3
70. GPT-5.4: 29.3
71. Qwen3.8-27B: 29.1
72. GPT-5.3 Codex: 29.1
73. GPT-6 Luna: 29
74. GLM-5.1: 29
75. Kimi K2.7 Code: 28.5
76. DeepSeek-V4-Flash-Max: 28.3
77. Inkling-Small: 28.3
78. Laguna S 2.1: 28.2
79. MiMo-V2.5: 27.6
80. Claude Opus 4.5: 27.2
81. Muse Spark 1.2: 27.1
82. MiniMax M2.7: 26.9
83. Qwen3.6 Plus: 26.7
84. Gemini 3.6 Flash: 26.6
85. MiniMax M2.5: 26.5
86. DeepSeek-V4-Flash-0423: 26.4
87. GPT-5: 26.2
88. Inkling: 26
89. GLM-5: 25.4
90. Qwen3.6-27B: 24.8
91. GPT-5.2 Codex: 24.5
92. Kimi K2.5: 24.4
93. MiMo-V2-Pro: 24.4
94. Claude Sonnet 4.6: 24.2
95. Ling 3.0 Flash: 24.1
96. Mistral Medium 3.5: 23.6
97. GPT-5.2: 23
98. DeepSeek-V3.2: 22.7
99. Muse Glimmer-30B: 22.7
100. Muse Spark: 22.7
101. Qwen3.5-397B-A17B: 22.7
102. DeepSeek-V3.2 (Thinking): 22.6
103. Seed 2.0 Pro: 22
104. MiniMax M2.1: 21.9
105. Gemini 3 Flash: 21.8
106. Claude Opus 4.1: 21.4
107. GPT-5.1 Instant: 21.3
108. GPT-5.4 mini: 21.2
109. GPT-5.1: 21.1
110. Nemotron 3.5 Super VL (120B A12B): 21.1
111. Claude Sonnet 4.5: 20.9
112. K2 Horizon 375B-A23B: 20.9
113. MiniCPM-SALA: 20.8
114. Kimi K2-Thinking-0905: 20.5
115. Solar Pro 4: 20.4
116. Step-3.5-Flash: 20.3
117. Qwen3.5-122B-A10B: 20.1
118. Nemotron 3 Ultra (550B A55B): 19.8
119. Claude Haiku 4.5: 19.5
120. Kimi K2 0905: 19.3
121. Nova 2 Pro: 19.2
122. MAI-Code-1-Flash: 18.9
123. LongCat-Flash-Thinking-2601: 18.8
124. Qwen3.6-35B-A3B: 18.8
125. Grok-3 Mini: 18.6
126. MiniMax M2: 18.1
127. Grok 4 Fast: 18
128. Solar Mini 4: 18
129. Grok-4 Heavy: 17.6
130. Grok-3: 17.5
131. Claude Opus 4: 17.4
132. GLM-4.7: 17.3
133. MiMo-V2-Flash: 17.3
134. DeepSeek-V3.2-Exp: 17.2
135. Grok-4: 17.2
136. Gemini 2.5 Pro Preview 06-05: 16.7
137. Qwen3.5-27B: 16.7
138. Claude Sonnet 4: 16.6
139. GLM-4.5: 15.9
140. Gemini 3.5 Flash-Lite: 15.5
141. GPT-5.4 nano: 15.4
142. GLM-4.6: 14.8
143. Nemotron 3.5 Bolt (74B A9B 24GB): 14.8
144. Nemotron Nano 9B v2: 14.3
145. o1-mini: 14.2
146. K2 Horizon MoVA 36B-A4B: 13.8
147. LongCat-Flash-Thinking: 13.7
148. Claude 3.7 Sonnet: 13.5
149. Nemotron 3 Super (120B A12B): 13
150. K2 Horizon 32B: 12.9
151. Mistral Large 2: 12.9
152. Grok Code Fast 1: 12.8
153. Claude 3.5 Sonnet: 12.8
154. Mercury 2: 12.7
155. Gemini 2.5 Pro: 12.6
156. DeepSeek-V3.1: 12.4
157. Llama 3.1 Nemotron Ultra 253B v1: 12.3
158. Qwen2.5 VL 32B Instruct: 12.2
159. LongCat-Flash-Chat: 12.2
160. Kimi K2 Instruct: 12.1
161. Qwen3 235B A22B: 12.1
162. Qwen3-Coder 480B A35B Instruct: 12.1
163. GPT-4o: 11.6
164. Qwen3.5-35B-A3B: 11.5
165. Qwen3 Max: 11.5
166. Llama 3.1 405B Instruct: 11.3
167. GLM-4.5-Air: 11.1
168. Claude 3.5 Sonnet: 11.1
169. Nova 2 Lite: 11.1
170. Granite 3.3 8B Instruct: 11
171. Granite 3.3 8B Base: 10.9
172. North Mini Code 1.0: 10.8
173. Mistral Small 4: 10.6
174. Nova Pro: 10
175. Qwen2.5-Coder 32B Instruct: 9.8
176. Qwen3 30B A3B: 9.7
177. Kimi K2-Instruct-0905: 9.6
178. Mistral Small 3.1 24B Instruct: 9.5
179. Qwen2.5 32B Instruct: 9.2
180. MiniMax M1 80K: 9.2
181. Llama 3.3 70B Instruct: 9.1
182. Grok-2: 9
183. Qwen3 32B: 8.8
184. IBM Granite 4.2 30B: 8.4
185. Gemini 2.5 Flash: 8.2
186. DeepSeek-R1-0528: 7.8
187. Command A+: 7.6
188. GPT-4 Turbo: 7.1
189. LongCat-Flash-Lite: 6.9
190. Qwen2.5-Coder 7B Instruct: 6.3
191. DeepSeek-V2.5: 5.9
192. Sarvam-30B: 5.4
193. GPT-4.5: 5.3
194. Gemma 3 27B: 5.2
195. GPT-4.1: 4.7
196. o1: 4
197. Qwen3-235B-A22B-Instruct-2507: 4
198. Claude 3.5 Haiku: 3.6
199. IBM Granite 4.2 8B: 3.5
200. Gemini Diffusion: 2.4
201. Mistral Large 3 (675B Instruct 2512): 1.4
202. Nemotron 3 Nano (30B A3B): 0.9
203. GPT-4o mini: -1
204. GPT-4o: -1.1
205. GPT-4.1 mini: -3.9
206. GPT-3.5 Turbo: -4.6
207. GPT-4.1 nano: -12.2

### SWE-rebench ⭐

Respected method, but its last window predates most current models.

Fresh GitHub issues, decontaminated — but the last window ends on 2026-07-01, before most current models. (updated 2026-07-24)

1. Fable 5 [high]: 64.5045
2. Grok 4.5 [high]: 63.7838
3. Opus 5 [high]: 63.4234
4. GLM-5.2 [high]: 62.8829
5. GPT-5.6 Sol [medium]: 62.3423
6. Junie: 61.8018
7. Claude Code: 60.3604
8. Codex: 58.018
9. Sonnet 5 [high]: 56.7568
10. Cursor: 51.7117
11. MiniMax M3: 47.2072
12. MiMo V2.5 Pro: 46.4865
13. GPT-5.6 Luna [medium]: 43.6036
14. DeepSeek-V4 Pro [high]: 40.1802
15. Qwen3.6-27B: 31.1712
16. Qwen3.6-35B-A3B: 24.6847
17. Qwen3.5-35B-A3B: 17.1171

### KiloBench (Terminal-Bench 2.0 in Kilo) ⭐

Runs the older Terminal-Bench 2.0 in a single agent, and is not refreshed often.

Terminal-Bench 2.0 run in the Kilo agent; cost is per full run of the suite, not per task. (updated 2026-09-28)

1. OpenAI: GPT-6 Astra ($$$$): 79.3258
2. OpenAI: GPT-5.6 Sol: 76.1798
3. Anthropic: Claude Fable 5.1 ($$$$): 76.1798
4. DeepSeek: DeepSeek V4.1 Flash: 75.2809
5. Google: Gemini 3.8 Flash (50% off): 75.2809
6. OpenAI: GPT-5.5: 74.1573
7. SpaceXAI: Grok 4.6: 73.0337
8. MoonshotAI: Kimi K3: 72.809
9. Anthropic: Claude Opus 5: 71.4607
10. Anthropic: Claude Fable 5 ($$$$): 71.0112
11. SpaceXAI: Grok 4.5: 70.7865
12. Anthropic: Claude Opus 4.7: 70.1124
13. Anthropic: Claude Opus 4.8: 67.6404
14. Google: Gemini 3.5 Flash: 64.7191
15. MoonshotAI: Kimi K2.7 Code: 60.6742
16. Meta: Muse Spark 1.1: 59.7753
17. Anthropic: Claude Sonnet 5: 59.5506
18. Anthropic: Claude Sonnet 4.6: 55.0562
19. Qwen: Qwen3.7 Max: 54.6067
20. MoonshotAI: Kimi K2.6: 54.382
21. Qwen: Qwen3.8 Max: 54.1573
22. Z.ai: GLM 5.2: 53.0337
23. Meta: Muse Spark 1.2: 51.4607
24. SpaceXAI: Grok Build 0.1: 50.5618
25. Kwaipilot: KAT-Coder-Pro V2.5: 50.3371
26. Z.ai: GLM 5.1: 49.4382
27. Google: Gemini 3.6 Flash: 47.8652
28. Tencent: Hy3 (free): 47.6404
29. Xiaomi: MiMo-V2.5-Pro: 47.6404
30. MiniMax: MiniMax M3: 47.6404
31. kilo-auto/efficient: 46.7416
32. DeepSeek: DeepSeek V4 Pro 0423: 44.0449
33. Thinking Machines: Inkling: 43.5955
34. Poolside: Laguna S 2.1: 31.0112
35. inclusionAI: Ling-2.6-1T (retires Aug 24): 28.0742
36. Poolside: Laguna XS 2.1: 26.7416
37. NVIDIA: Nemotron 3 Ultra: 19.1011
38. NVIDIA: Nemotron 3 Super (free): 15.5056

### Terminal-Bench-Science 0.1 ⭐

Version 0.1 mirror, superseded by the Artificial Analysis run of the same benchmark.

Scientific work in a terminal (life, physical, earth, engineering and mathematical sciences) done by coding agents, each model in its vendor's agent at a stated effort. Official board, rows from Snorkel's mirror. (updated 2026-09-24)

1. GPT-6 Astra (Codex, max): 68.1 ±3.2
2. Opus 5.5 (Claude Code, max): 63.3 ±3.3
3. Fable 5.1 (Claude Code, max): 40 ±3.4
4. Opus 5 (Claude Code, max): 30 ±3.2
5. GPT-5.6 Sol (Codex, max): 22.4 ±2.9
6. Fable 5 (Claude Code, max): 21.4 ±2.8
7. DeepSeek V4.1 Flash (Codex, max): 15.7 ±2.5
8. Grok 4.7 (Grok Build, xhigh): 14.3 ±2.4
9. Gemini 3.8 Flash (mini-SWE-agent, high): 12.4 ±2.3
10. Opus 4.8 (Claude Code, max): 10.5 ±2.1
11. GPT-5.6 Terra (Codex, max): 8.6 ±1.9
12. GLM 5.3 (Claude Code, max): 8.1 ±1.9
13. Kimi K3 (Claude Code, max): 7.1 ±1.8
14. Grok 4.6 (Grok Build, xhigh): 7.1 ±1.8
15. Gemini 3.7 Flash (mini-SWE-agent, high): 5.7 ±1.6
16. DeepSeek V4 Pro (Codex, max): 3.8 ±1.3
17. GPT-5.6 Luna (Codex, max): 3.3 ±1.2

### Snorkel LibraryDesignBench ⭐

Narrow (library API design), few models, not refreshed recently.

Can the model design a library that other code then uses? Pass rate on downstream problems times the simplicity of the API, per model and harness. Run by Snorkel AI. (updated 2026-10-01)

1. Opus 5.5 (mini-SWE): 48.9 ±0.9
2. Fable 5.1 (mini-SWE): 47.5 ±1.1
3. GPT-6 Astra (mini-SWE): 45.1 ±0.7
4. GPT-6 Astra (Codex): 44.1 ±0.8
5. Kimi K3 (mini-SWE): 44 ±1
6. Fable 5.1 (Claude Code): 42.6 ±2.5
7. GLM 5.3 (mini-SWE): 41.9 ±1.1
8. Grok 4.6 (mini-SWE): 39.7 ±1.1
9. GPT-5.6 Sol (Codex): 39.5 ±1
10. GPT-6 Sol (mini-SWE): 38.5 ±0.7
11. DeepSeek V4 Pro (mini-SWE): 31.2 ±0.8

## Official launch figures

What each vendor published for its own model at launch, with the effort it used. Vendors pick their own agent, harness and settings, so these numbers usually run higher than independent measurements — use them to see each vendor's claims, and the charts above to compare.

- Claude Fable 5.2 · max · SWE-bench Pro: 93. [Claude Fable 5.2 and Claude Mythos 5.2 System Card](https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card) (2026-11-11)
- Claude Fable 5.2 · max · Terminal-Bench 4.0: 67.6. [Claude Fable 5.2 and Claude Mythos 5.2 System Card](https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card) (2026-11-11)
- Claude Fable 5.2 · max · DeepSWE v1.1: 74.3. [Claude Fable 5.2 and Claude Mythos 5.2 System Card](https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card) (2026-11-11)
- Claude Fable 5.2 · xhigh · FrontierCode 2.0: 49.1. [Introducing Claude Fable 5.2 and Claude Mythos 5.2](https://www.anthropic.com/claude-fable-and-mythos-5-2) (2026-11-11) · [Claude Fable 5.2 and Claude Mythos 5.2 System Card](https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card) (2026-11-11)
- Claude Opus 5.5 · max · SWE-bench Pro: 89.9. [Claude Opus 5.5 System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf) (2026-09-28)
- Claude Opus 5.5 · xhigh · Terminal-Bench 4.0: 66.4. [Claude Opus 5.5 System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf) (2026-09-28)
- Claude Opus 5.5 · max · DeepSWE v1.1: 74.2. [Claude Opus 5.5 System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf) (2026-09-28)
- Claude Opus 5.5 · max · FrontierCode 1.1: 54.4. [Claude Opus 5.5 System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf) (2026-09-28)
- Claude Fable 5.1 · max · SWE-bench Pro: 81.2. [Claude Opus 5.5 System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf) (2026-09-28)
- Claude Fable 5.1 · max · Terminal-Bench 4.0: 55.8. [Claude Opus 5.5 System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf) (2026-09-28)
- Claude Fable 5.1 · max · FrontierCode 1.1: 50.3. [Claude Opus 5.5 System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf) (2026-09-28)
- Claude Sonnet 5.5 · max · SWE-bench Pro: 81.3. [Introducing Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5) (2026-09-29)
- Claude Sonnet 5.5 · max · Terminal-Bench 4.0: 70.6. [Introducing Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5) (2026-09-29)
- Claude Sonnet 5.5 · max · DeepSWE v1.1: 71. [Claude Sonnet 5.5 System Card](https://www.anthropic.com/claude-sonnet-5-5-system-card) (2026-09-29)
- Claude Sonnet 5.5 · xhigh · FrontierCode 1.1: 52.1. [Introducing Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5) (2026-09-29)
- Claude Haiku 5.5 · max · SWE-bench Pro: 64.8. [Claude Haiku 5.5 System Card](https://www.anthropic.com/claude-haiku-5-5-system-card) (2026-10-10)
- Claude Haiku 5.5 · max · Terminal-Bench 4.0: 39.2. [Introducing Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5) (2026-10-10)
- Claude Haiku 5.5 · max · FrontierCode 1.1: 46.4. [Introducing Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5) (2026-10-10)
- Claude Sonnet 5 · max · SWE-bench Verified: 85.2. [Claude Sonnet 5 System Card](https://www-cdn.anthropic.com/480e0bb54327b9622282e9c39a83a4f490ed377e/Claude%20Sonnet%205%20System%20Card.pdf) (2026-09-28)
- Claude Sonnet 5 · max · SWE-bench Pro: 63.2. [Claude Sonnet 5 System Card](https://www-cdn.anthropic.com/480e0bb54327b9622282e9c39a83a4f490ed377e/Claude%20Sonnet%205%20System%20Card.pdf) (2026-09-28)
- Claude Sonnet 5 · xhigh · Terminal-Bench 2.x: 80.4 (2.1). [Claude Sonnet 5 System Card](https://www-cdn.anthropic.com/480e0bb54327b9622282e9c39a83a4f490ed377e/Claude%20Sonnet%205%20System%20Card.pdf) (2026-09-28)
- Claude Haiku 4.5 · — · SWE-bench Verified: 73.3. [Introducing Claude Haiku 4.5](https://www.anthropic.com/news/claude-haiku-4-5) (2026-09-28)
- GPT-6 Astra · high · Terminal-Bench 4.0: 57.88. [GPT-6 Astra](https://openai.com/index/gpt-6-astra/) (2026-09-28)
- GPT-6 Astra · xhigh · DeepSWE v1.1: 74.12. [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) (2026-10-10)
- GPT-6 Astra · max · FrontierCode 1.1: 53.3. [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) (2026-10-10)
- GPT-6.1 Sol · high · DeepSWE v1.1: 75.22. [Introducing GPT-6.1 Sol](https://openai.com/index/introducing-gpt-6-1-sol/) (2026-10-10)
- GPT-6.1 Terra · max · Terminal-Bench 4.0: 50.3. [Introducing GPT-6.1 Terra](https://openai.com/index/introducing-gpt-6-1-terra/) (2026-10-27)
- GPT-6 Sol · max · DeepSWE v1.1: 68.81. [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) (2026-10-10)
- GPT-6 Sol · max · FrontierCode 1.1: 49.27. [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) (2026-10-10)
- GPT-6 Luna · max · DeepSWE v1.1: 66.59. [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) (2026-10-10)
- GPT-6 Luna · max · FrontierCode 1.1: 42.42. [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) (2026-10-10)
- Gemini 4 Carbon · high · Terminal-Bench 4.0: 65.2. [Gemini 4 Carbon model evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_4_carbon_model_evaluation.pdf) (2026-11-11) · [Gemini 4 Carbon model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Carbon-Model-Card.pdf) (2026-11-11)
- Gemini 4 Argon · high · DeepSWE v1.1: 77.9. [Gemini 4 Argon model evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_4_argon_model_evaluation.pdf) (2026-10-10) · [Gemini 4 Argon model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Argon-Model-Card.pdf) (2026-10-17)
- Gemini 4 Argon · high · Terminal-Bench 4.0: 57.4. [Gemini 4 Argon model evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_4_argon_model_evaluation.pdf) (2026-10-10) · [Gemini 4 Argon model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Argon-Model-Card.pdf) (2026-10-17)
- Gemini 3.8 Flash · medium · Terminal-Bench 2.x: 89.4 (2.1). [Gemini 3.8 Flash evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_3-8_flash_model_evaluation.pdf) (2026-09-28)
- Gemini 3.8 Flash · — · Terminal-Bench 4.0: 19.1. [Gemini 3.8 Flash evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_3-8_flash_model_evaluation.pdf) (2026-09-28)
- Gemini 3.8 Flash · high · DeepSWE v1.1: 73.7. [Gemini 3.8 Flash evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_3-8_flash_model_evaluation.pdf) (2026-09-28)
- Gemini 3.7 Flash (deprecated) · medium · Terminal-Bench 2.x: 85.8 (2.1). [Gemini 3.8 Flash evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_3-8_flash_model_evaluation.pdf) (2026-09-28)
- Gemini 3.7 Flash (deprecated) · — · Terminal-Bench 4.0: 11.2. [Gemini 3.8 Flash evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_3-8_flash_model_evaluation.pdf) (2026-09-28)
- Gemini 3.7 Flash (deprecated) · — · DeepSWE v1.1: 65.3. [Gemini 3.8 Flash evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_3-8_flash_model_evaluation.pdf) (2026-09-28)
- Gemini 3.1 Pro (preview) · high · SWE-bench Verified: 80.6. [Gemini 3.1 Pro model card](https://deepmind.google/models/model-cards/gemini-3-1-pro/) (2026-09-28)
- Gemini 3.1 Pro (preview) · high · SWE-bench Pro: 54.2. [Gemini 3.1 Pro model card](https://deepmind.google/models/model-cards/gemini-3-1-pro/) (2026-09-28)
- Gemini 3.1 Pro (preview) · high · Terminal-Bench 2.x: 68.5 (2.0). [Gemini 3.1 Pro model card](https://deepmind.google/models/model-cards/gemini-3-1-pro/) (2026-09-28)
- Grok 4.7 · xhigh · Terminal-Bench 4.0: 38. [Grok 4.7 model card](https://media.x.ai/v1/website/4p7card-5eccc980.pdf) (2026-09-28)
- Grok 4.7 · high · DeepSWE v1.1: 71. [Grok 4.7](https://x.ai/news/grok-4-7) (2026-09-28)
- Grok 4.6 · high · Terminal-Bench 4.0: 20.3. [Grok 4.7 model card](https://media.x.ai/v1/website/4p7card-5eccc980.pdf) (2026-09-28)
- Grok 4.6 · high · DeepSWE v1.1: 65.9. [Grok 4.6](https://x.ai/news/grok-4-6) (2026-09-28)
- Kimi K3 · max · Terminal-Bench 2.x: 88.3 (2.1). [Kimi K3](https://github.com/MoonshotAI/Kimi-K3) (2026-09-28)
- Kimi K3 · max · DeepSWE v1.1: 67.5. [Kimi K3](https://github.com/MoonshotAI/Kimi-K3) (2026-09-28)
- Kimi K3.1 Pro · max · Terminal-Bench 4.0: 65.7. [Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence](https://www.kimi.ai/blog/kimi-k3-1-pro) (2026-10-27)
- Kimi K3.1 Pro · max · DeepSWE v1.1: 75.5. [Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence](https://www.kimi.ai/blog/kimi-k3-1-pro) (2026-10-27)
- GLM-5.3 · max · Terminal-Bench 2.x: 88.2 (2.1). [GLM-5.3 model card](https://huggingface.co/zai-org/GLM-5.3) (2026-09-28)
- GLM-5.3 · max · DeepSWE v1.1: 66.9. [GLM-5.3 model card](https://huggingface.co/zai-org/GLM-5.3) (2026-09-28)
- GLM-5.3-Flash · — · Terminal-Bench 2.x: 84.3 (2.1). [GLM-5.3-Flash](https://z.ai/blog/glm-5.3-flash) (2026-09-28)
- GLM-5.3-Flash · — · DeepSWE v1.1: 63.4. [GLM-5.3-Flash](https://z.ai/blog/glm-5.3-flash) (2026-09-28)
- Mistral Large 4 · — · Terminal-Bench 4.0: 41.4. [Mistral Large 4 (Le Chonk): Built for Coders, Trusted by Defenders](https://mistral.ai/news/mistral-large-4-oss/) (2026-11-11)
- Mistral Large 4 · high · FrontierCode 2.0: 19.7. [Mistral Large 4 (Le Chonk): Built for Coders, Trusted by Defenders](https://mistral.ai/news/mistral-large-4-oss/) (2026-11-11)
- Mistral Large 4 · — · DeepSWE v1.1: 61.7 (preview). [Introducing Mistral Large 4](https://mistral.ai/news/mistral-large-4/) (2026-10-10)
## Remote or local?

When running a model on your own machine makes sense: hardware, performance and limits compared with remote agents.


- **24 GB GPU (RTX 4090 / 3090)**: Qwen3.8-27B Q4_K_M, 16.5 GB, 38–46 tok/s, AA 33.7 · TB4 5.6%
- **32 GB GPU (RTX 5090)**: Qwen3.8-27B Q6 / NVFP4, 22–25 GB, 23–75 tok/s · MTP 98–134, AA 33.7 · TB4 5.6%
- **128 GB unified memory (Strix Halo, DGX Spark, M5 Max)**: Qwen3.8-Flash-Next IQ4_XS, 93.7 GB, 11–37 tok/s · MTP 47–84, AA 39.8 · TB4 25.3%
- **256 GB unified memory (M5 Ultra, M3 Ultra)**: GLM-5.3-Flash MLX 4-bit, 204 GB, 26–41 tok/s, AA 41.8 · TB4 32.8%

## FAQ

Focused answers to choose a model, balance reasoning effort and manage subscription quotas.

### What is an AI coding agent?

[Link to this question](https://coding-agent.josedacosta.net/en#q-what-is-a-coding-agent)

An AI that works inside a software project: it reads the files, edits the code, runs commands and tests, and iterates until the task is done, under the developer's control. The most used in 2026 are Claude Code, GitHub Copilot, Codex and Cursor.

Where and how an agent is used: a terminal (CLI), a desktop app, an IDE extension, a hosted web service, remotely from a phone or another browser, inside GitHub and CI/CD, from team tools (Slack, Teams, Linear, Jira) or through an SDK in your own scripts. The same agent often has several, each with its own name: Codex CLI, the ChatGPT desktop app, the Codex IDE extension, Codex cloud, Codex Remote, Codex code review and the Codex SDK.

Sources: [AI coding agent adoption 2026](https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/) · [Codex glossary](https://learn.chatgpt.com/codex/glossary)

### Which agent, model and effort for “Advanced planning and architecture”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-planning)

Claude Code · Claude Fable 5.2 · high, with Claude Max ($100) or Ultra ($200): Fable may use up to 50% of the weekly limit (on Pro, paid credits only). ≈ 2.9 agent-hours per 5-hour window · ≈ 14 per week on Claude Max 5x (central estimate). Or: Claude Code · Claude Opus 5.5 (default) · xhigh, with Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 1.7 agent-hours per 5-hour window · ≈ 13 per week on Claude Pro (central estimate).

How long: Only for the plan: one session, until the plan or `SPEC.md` is written, reviewed and approved. Then: Stop there. Implement in a fresh session with Opus 5.5 at medium (default) — see Everyday coding — or hand the plan to a long run with `/goal`. Staying at xhigh to implement costs 2.5× more per task (AA index: $6.05 vs $2.45), and implementing with Fable 5.2 would spend its weekly cap on the part a cheaper model does well.

Why: A plan is one long reasoning pass, so effort pays off: Opus 5.5 scores 57.4 on the AA Intelligence Index (v5) at xhigh, above Claude Sonnet 5.5 at max (54.5 for $10.02) and just below its own max (59.1), for 40% less than max ($6.05 vs $10.04 per task). It works on Claude Pro.

Sources: [Artificial Analysis model leaderboard](https://artificialanalysis.ai/leaderboards/models) · [Snorkel LibraryDesignBench 1.0.0](https://snorkel.ai/leaderboard/librarydesignbench/) · [CWE-bench v2](https://cwe-bench.com/) · [LiveBench](https://livebench.ai/)

### Which agent, model and effort for “Everyday coding”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-daily)

Claude Code · Claude Opus 5.5 (default) · medium (default), with Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).

How long: All day: medium, Claude Code's default, is the setting to leave on. One session per task (`/clear` between tasks), and end a session after two failed corrections. Then: A task that resists (a subtle change, two failed attempts): `/effort high` for that task only, or `ultrathink` for one turn, then back to medium. A bug you cannot explain: Debugging. A change you cannot describe in one sentence: plan it first (Advanced planning). Many small mechanical edits: High volume.

Why: The best quality for the quota on routine work. Benchmarks are made of hard tasks: there, high adds about 4 points over medium (CursorBench 4.0 52.5% → 56.0%, Terminal-Bench 4.0 (AA) 52.5% → 56.6%, FrontierCode 2.0 30.1% → 34.5%) for about 35% more quota, so about a quarter fewer agent-hours. On everyday tasks that medium already solves, that gain does not show; keep high for the tasks that resist.

Sources: [Artificial Analysis model leaderboard](https://artificialanalysis.ai/leaderboards/models) · [FrontierCode leaderboard data](https://cognition.com/data/frontiercode-leaderboard/data.json) · [Arena Code: WebDev leaderboard](https://arena.ai/leaderboard/code/webdev) · [Design Arena: Code overall](https://www.designarena.ai/leaderboard/code)

### Which agent, model and effort for “Debugging”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-debug)

Claude Code · Claude Opus 5.5 (default) · high, with Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 1.8 agent-hours per 5-hour window · ≈ 14 per week on Claude Pro (central estimate).

How long: Until the cause is found and the fix is proven by a test, then `/effort medium` again: high costs about 27% more per task than medium, too much to leave on all day. Then: Still stuck: add `ultrathink` to the next prompt (one turn at xhigh). Stuck after two attempts: `/clear`, then Claude Fable 5.2 at xhigh (Max or Ultra) with what you learned. Once fixed: back to Everyday coding.

Why: Diagnosis is a hard task, where effort shows. Terminal-Bench is mostly diagnosis in a terminal (failing builds, broken services, data to recover); on the Artificial Analysis run, Opus 5.5 gains 4 points from medium to high (52.5% → 56.6%) for $4.04 → $5.12 per task, about 27% more quota.

Sources: [Artificial Analysis model leaderboard](https://artificialanalysis.ai/leaderboards/models) · [CursorBench 4.0](https://cursor.com/cursorbench) · [Vals AI benchmarks](https://www.vals.ai/benchmarks) · [Terminal-Bench 4.1 leaderboard](https://www.tbench.ai/)

### Which agent, model and effort for “Front-end and web UI”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-web)

Claude Code · Claude Opus 5.5 (default) · medium (default), with Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).

How long: For pages, components and tweaks: medium, the default, already leads Design Arena, whose Opus 5.5 entries run at a default-like effort. Then: For the one page that must stand out (a landing page, a showcase): `/effort max` while you build it, the setting behind the Arena WebDev lead, then back to medium. Max burns about 35% more quota per hour (effort factor 1 vs 0.74).

Why: Top of Arena WebDev (1814.3 ±11.7, blind human votes, cutoff 2026-11-10), shared with Claude Fable 5.2 at xhigh (1835.8 ±22.4 on only 994 votes, ranks 1–2), and clearly ahead of Claude Sonnet 5.5 xhigh (1772.5), Gemini 4 Carbon high (1768.2, pre-release), GPT-6 Astra max (1764.4) and GPT-6.1 Sol max (1743.0). First on Arena Image-to-WebDev too (1751.9 ±19.2), level with Claude Sonnet 5.5 (1744.1).

Sources: [Arena Code: WebDev leaderboard](https://arena.ai/leaderboard/code/webdev) · [Arena Code: Image-to-WebDev](https://arena.ai/leaderboard/code/image-to-webdev) · [Design Arena: Code overall](https://www.designarena.ai/leaderboard/code) · [Design Arena: Website](https://www.designarena.ai/leaderboard/website)

### Which agent, model and effort for “Long autonomous agent runs”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-long-runs)

Claude Code · Claude Opus 5.5 (default) · xhigh, with Claude Max ($100) or Ultra ($200) for long sessions. One agent never hits the 5-hour limit (≈ 8.6 agent-hours per window) · ≈ 82.2 agent-hours per week on Claude Max 5x (central estimate). Or: Codex · GPT-6.1 Sol (default) · xhigh, with ChatGPT Plus ($20) or Pro. ≈ 3.7 agent-hours per 5-hour window · ≈ 26.3 per week on ChatGPT Plus (Codex) (central estimate).

How long: For the whole run, until the `/goal` condition is met — hours if needed. Check the milestones instead of watching every step. Then: When the goal is met, end the run, review in a fresh context (`/code-review`), then go back to medium (Everyday coding) for follow-ups. If the run fails on the same error twice, stop it and re-plan instead of letting it burn quota.

Why: On the AA Coding Agent Index, which runs each model in its vendor's agent, Opus 5.5 at max in Claude Code is first (61.7 for $11.42 per task), ahead of Gemini 4 Carbon (59.4) and Claude Sonnet 5.5 (59.3); GPT-6.1 Sol at max in Codex scores 54.1 for $1.50, lower but for about an eighth of the cost. On the official tbench.ai leaderboard (Terminal-Bench 4.1, the 2026-11-09 regrade of 4.0), Opus 5.5 at max is first (65.76%, $14.30 per trial) and GPT-6.1 Sol at max (60.91%) edges GPT-6 Astra (60.30%) for a fifth of its cost per trial ($1.92 vs $9.90).

Sources: [Artificial Analysis model leaderboard](https://artificialanalysis.ai/leaderboards/models) · [Terminal-Bench 4.1 leaderboard](https://www.tbench.ai/) · [Artificial Analysis Coding Agent Index](https://artificialanalysis.ai/agents/coding-agents) · [Introducing Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5)

### Which agent, model and effort for “High volume, subagents, quick edits”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-volume)

Codex · GPT-6.1 Terra · medium (default), with ChatGPT Plus ($20). One agent never hits the 5-hour limit (≈ 6.6 agent-hours per window) · ≈ 47.4 agent-hours per week on ChatGPT Plus (Codex) (central estimate). Or: Claude Code · Claude Haiku 5.5 · medium (default), with Claude Pro ($20), Max ($100) or Ultra ($200). One agent never hits the 5-hour limit (≈ 16.1 agent-hours per window) · ≈ 122.3 agent-hours per week on Claude Pro (central estimate).

How long: As long as the tasks stay mechanical and fully specified: renames, types, docstrings, tests from a clear pattern. Test the prompt on 2–3 files, then fan out. Then: As soon as a task needs judgment (a design choice, an unclear bug), go back to Everyday coding. Only when the Terra quota runs out on trivial edits, GPT-6 Luna at max costs a little less, but scores far lower (Terminal-Bench 4.0 (AA) 12.6% against 32.3%). Never Claude Fable 5.2 for bulk work or subagents: it would empty its 50% weekly cap on edits a small model does as well.

Why: Quality first, then price, for easy tasks: bulk edits are mechanical and fully specified, so the default effort is enough. At medium, GPT-6.1 Terra scores 32.3% on Terminal-Bench 4.0 (AA) for $0.28 per task, 38.2 on the AA index for $0.17 and 42.0 on the AA Coding Agent Index for $0.39: ahead of Claude Haiku 5.5 at medium (15.2%, 24.6) and far ahead of GPT-6 Luna at max (12.6%, 27.9, 31.2).

Sources: [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) · [Introducing GPT-6.1 Sol](https://openai.com/index/introducing-gpt-6-1-sol/) · [Introducing GPT-6.1 Terra](https://openai.com/index/introducing-gpt-6-1-terra/) · [Codex pricing](https://learn.chatgpt.com/docs/pricing)

### Which agent, model and effort for “Algorithms, proofs and competitive programming”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-math)

Codex · GPT-6 Astra · max, with ChatGPT Pro ($100 / $200 / $500). ≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate). Or: Codex · GPT-6.1 Sol (default) · max, with ChatGPT Plus ($20) or Pro. ≈ 3.7 agent-hours per 5-hour window · ≈ 26.3 per week on ChatGPT Plus (Codex) (central estimate).

How long: Only for the hard core — the algorithm, the proof, the optimization: one focused session. Then: Once the core is correct and tested, go back to Everyday coding for the surrounding code, I/O and tests.

Why: Quality first: GPT-6 Astra at max leads ALE-Bench (2951, against 2147 for Opus 5.5 at high), scores 100% on Vals IOI (Opus 5.5 95.1% ±4.9) and leads WeirdML v3 (42.2 at xhigh, Opus 5.5 31.2).

Sources: [Vals AI benchmarks](https://www.vals.ai/benchmarks) · [ALE-Bench (Epoch AI Benchmarking Hub)](https://epoch.ai/data/benchmark_data.zip) · [WeirdML v3 (Epoch AI Benchmarking Hub)](https://epoch.ai/data/benchmark_data.zip) · [LiveBench](https://livebench.ai/)

### Which agent, model and effort for “Scientific and numerical code”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-science)

Claude Code · Claude Fable 5.2 · high, with Claude Max ($100) or Ultra ($200): Fable may use up to 50% of the weekly limit (on Pro, paid credits only). ≈ 2.9 agent-hours per 5-hour window · ≈ 14 per week on Claude Max 5x (central estimate). Or: Claude Code · Claude Opus 5.5 (default) · xhigh, with Claude Max ($100) or Ultra ($200) for long sessions. One agent never hits the 5-hour limit (≈ 8.6 agent-hours per window) · ≈ 82.2 agent-hours per week on Claude Max 5x (central estimate).

How long: For the numerical core: the model, the formulas, the kernel, until it matches a reference result. Then: Once the core matches its reference, go back to Everyday coding (Opus 5.5 at medium) for I/O, plots and packaging: Fable 5.2 would spend its weekly cap on code a cheaper model writes as well.

Why: Quality first: Claude Fable 5.2 leads Terminal-Bench Science (AA) at high (70.0%; 73.3% at max), well ahead of Gemini 4 Carbon (64.8%), GPT-6 Astra (63.3% at max) and Opus 5.5 at xhigh (61.9%). High rather than xhigh (67.6%) or max, which costs far more for 3 points.

Sources: [Artificial Analysis evaluation details](https://artificialanalysis.ai/evaluations/terminalbench-4-0) · [Artificial Analysis: Terminal-Bench-Science](https://artificialanalysis.ai/evaluations/terminal-bench-science) · [Terminal-Bench-Science 0.1 (Snorkel mirror of the official board)](https://snorkel.ai/leaderboard/terminal-bench-science/) · [Artificial Analysis model leaderboard](https://artificialanalysis.ai/leaderboards/models)

### Which agent, model and effort for “SVG illustrations and icons as code”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-svg-diagrams)

Codex · GPT-6 Astra · xhigh, with ChatGPT Plus ($20, 5–45 Astra messages per 5 hours) or Pro ($100 / $200 / $500). ≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate). Or: Claude Code · Claude Opus 5.5 (default) · high, with Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 1.8 agent-hours per 5-hour window · ≈ 14 per week on Claude Pro (central estimate).

How long: While you create the assets — an icon set, an illustration, a logo — iterating with renders until it looks right. Then: For small edits to an existing SVG (colors, sizes), your everyday agent is enough.

Why: First general-purpose model in the Design Arena SVG arena: Elo 1421 ±9 over 1,808 blind battles (68.8% won), about 90 points above Gemini 4 Argon (1329), Kimi K3.1 Pro (1326) and Claude Opus 5.5 (1321 ±10, 1,261 battles): about seven standard errors, the only large-sample signal for SVG illustrations. Only the SVG-specialized Arrow 2 Telos, in no subscription, scores higher.

Sources: [Design Arena: SVG](https://www.designarena.ai/leaderboard/svgs) · [MacBook SVG Benchmark: 34 AI Models, Ranked by the Laptop They Draw](https://playcode.io/blog/macbook-svg-benchmark) · [Managing usage with GPT-6 Astra in Work and Codex](https://help.openai.com/en/articles/20001516-managing-usage-with-gpt-6-astra-in-work-and-codex) · [Codex pricing](https://learn.chatgpt.com/docs/pricing)

### Which agent, model and effort for “Charts, dashboards and diagrams as code”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-charts-diagrams)

Claude Code · Claude Opus 5.5 (default) · medium (default), with Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).

How long: While you build the chart or the first version of the diagram, iterating with renders until it reads right. Then: For later edits (a label, a new node, a color), your everyday agent is enough: a Mermaid diagram is ordinary text once it exists.

Why: Claude Opus 5.5 leads Design Arena data visualization (1390 ±14 over 700 blind battles), ahead of Claude Sonnet 5.5 (1362, on only 160 battles), Kimi K3 (1348), Claude Fable 5.1 (1342), Gemini 4 Argon (1333) and GPT-6 Astra (1325, 10th): about three standard errors ahead of Astra, the leader for SVG illustrations.

Sources: [Design Arena: data visualization](https://www.designarena.ai/leaderboard) · [MacBook SVG Benchmark: 34 AI Models, Ranked by the Laptop They Draw](https://playcode.io/blog/macbook-svg-benchmark) · [Prompting Claude Opus 5.5](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5) · [Antigravity models](https://antigravity.google/docs/models/)

### Which agent, model and effort for “3D, games and creative coding as code”?

[Link to this question](https://coding-agent.josedacosta.net/en#q-use-case-3d)

Claude Code · Claude Opus 5.5 (default) · medium (default), with Claude Pro ($20), Max ($100) or Ultra ($200). ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate). Or: Codex · GPT-6 Astra · xhigh, with ChatGPT Pro ($100 / $200 / $500). ≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate).

How long: While you build the scene and its interactions, at the default effort, medium. Then: For tuning values, colors and small fixes, `/effort medium` (default). Textures and icons as SVG: see SVG illustrations and icons.

Why: Games: the best Claude model on Design Arena game dev (1429, 2,666 battles), behind GPT-6 Astra at max (1466) and xhigh (1457), and ahead of Claude Fable 5.2 (1404), on browser games written as code.

Sources: [Design Arena: Game dev](https://www.designarena.ai/leaderboard/game-dev) · [Arena Code: WebDev leaderboard](https://arena.ai/leaderboard/code/webdev) · [Design Arena: Code overall](https://www.designarena.ai/leaderboard/code) · [Managing usage with GPT-6 Astra in Work and Codex](https://help.openai.com/en/articles/20001516-managing-usage-with-gpt-6-astra-in-work-and-codex)

### Is max effort worth it?

[Link to this question](https://coding-agent.josedacosta.net/en#q-is-max-effort-worth-it)

Rarely. The agent's default effort for everyday work; high for debugging, xhigh for planning and multi-hour runs; max only where you have measured a gain.

Opus 5.5: on Terminal-Bench 4.0 (AA) max equals xhigh (59.6%) for 49% more. Claude Fable 5.2: xhigh beats max on FrontierCode 2.0 (49.1% vs 45.7%) for a third of the cost. GPT-6 Astra: xhigh matches or beats max on Terminal-Bench 4.0 and DeepSWE (59.6% vs 59.1%; 74.1% vs 73.2%) for about 30% less. Anthropic warns that max can overthink.

Sources: [FrontierCode leaderboard data](https://cognition.com/data/frontiercode-leaderboard/data.json) · [Artificial Analysis model leaderboard](https://artificialanalysis.ai/leaderboards/models) · [DeepSWE v1.1 leaderboard](https://deepswe.datacurve.ai/) · [Model configuration](https://code.claude.com/docs/en/model-config)

### How much do coding agent subscriptions cost?

[Link to this question](https://coding-agent.josedacosta.net/en#q-subscription-prices)

Monthly prices in US dollars, by agent:

- Claude Code: Claude Pro $20, Claude Max 5x $100, Claude Ultra $200
- Codex: ChatGPT Plus (Codex) $20, ChatGPT Pro 100 (Codex) $100, ChatGPT Pro 200 (Codex) $200, ChatGPT Pro 500 (Codex) $500
- Antigravity: Google AI Pro (Antigravity) $19.99, Google AI Ultra 5x (Antigravity) $99.99, Google AI Ultra 20x (Antigravity) $199.99
- Grok Build: X Subscription Premium (Grok Build) $30
- Claude Code / ZCode: GLM Coding Plan Lite $18, GLM Coding Plan Pro $80, GLM Coding Plan Max $168
- Qwen Code: Qwen Token Plan Lite $6, Qwen Token Plan Essential $10, Qwen Token Plan Standard $18, Qwen Token Plan Pro $68
- MiMo Code: MiMo Token Plan Lite $6, MiMo Token Plan Standard $16, MiMo Token Plan Pro $50, MiMo Token Plan Max $100
- GitHub Copilot: $10 / $39 / $100
- OpenCode Go: $10 / $40 / $20 per member
- Cursor: $8 / $30 / $100 / $200

Sources: [Claude pricing](https://claude.com/pricing) · [What is the Max plan?](https://support.claude.com/en/articles/11049741) · [What is the Ultra plan?](https://support.claude.com/en/articles/17728952) · [Codex pricing](https://learn.chatgpt.com/docs/pricing) · [About ChatGPT Pro tiers](https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers) · [Antigravity plans](https://antigravity.google/docs/plans/) · [Changes to Antigravity plans](https://antigravity.google/blog/changes-to-antigravity-plans) · [Grok FAQ](https://docs.x.ai/grok/faq)

### How do the 5-hour and weekly limits work?

[Link to this question](https://coding-agent.josedacosta.net/en#q-five-hour-and-weekly-limits)

A rolling limit counted over 5 hours from the first request, then reset. Claude and ChatGPT (Codex) plans combine it with a weekly limit.

A second cap counted over 7 days, which stops heavy users even when each 5-hour window still has room. On Claude Max and Ultra plans, the Claude Fable models (5.2, 5.1, 5) may use at most 50% of it together. Claude Ultra has only this weekly limit (no 5-hour window once the account is verified).

The simulator estimates how many agent-hours each subscription gives, and when the quota runs out.

Sources: [How do usage and length limits work?](https://support.claude.com/en/articles/11647753) · [Codex pricing](https://learn.chatgpt.com/docs/pricing) · [Claude Fable models on your plan](https://support.claude.com/en/articles/15424964)

### How many hours of agent work does a subscription give?

[Link to this question](https://coding-agent.josedacosta.net/en#q-agent-hours-per-subscription)

Central estimates of the simulator for the current picks (one agent working continuously):

- Advanced planning and architecture (Claude Code · Claude Fable 5.2 · high): ≈ 2.9 agent-hours per 5-hour window · ≈ 14 per week on Claude Max 5x (central estimate).
- Everyday coding (Claude Code · Claude Opus 5.5 (default) · medium (default)): ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).
- Debugging (Claude Code · Claude Opus 5.5 (default) · high): ≈ 1.8 agent-hours per 5-hour window · ≈ 14 per week on Claude Pro (central estimate).
- Front-end and web UI (Claude Code · Claude Opus 5.5 (default) · medium (default)): ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).
- Long autonomous agent runs (Claude Code · Claude Opus 5.5 (default) · xhigh): One agent never hits the 5-hour limit (≈ 8.6 agent-hours per window) · ≈ 82.2 agent-hours per week on Claude Max 5x (central estimate).
- High volume, subagents, quick edits (Codex · GPT-6.1 Terra · medium (default)): One agent never hits the 5-hour limit (≈ 6.6 agent-hours per window) · ≈ 47.4 agent-hours per week on ChatGPT Plus (Codex) (central estimate).
- Algorithms, proofs and competitive programming (Codex · GPT-6 Astra · max): ≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate).
- Scientific and numerical code (Claude Code · Claude Fable 5.2 · high): ≈ 2.9 agent-hours per 5-hour window · ≈ 14 per week on Claude Max 5x (central estimate).
- SVG illustrations and icons as code (Codex · GPT-6 Astra · xhigh): ≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate).
- Charts, dashboards and diagrams as code (Claude Code · Claude Opus 5.5 (default) · medium (default)): ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).
- 3D, games and creative coding as code (Claude Code · Claude Opus 5.5 (default) · medium (default)): ≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate).

Sources: [How do usage and length limits work?](https://support.claude.com/en/articles/11647753) · [Codex pricing](https://learn.chatgpt.com/docs/pricing)

### Which model leads the coding leaderboards right now?

[Link to this question](https://coding-agent.josedacosta.net/en#q-coding-leaderboard-leaders)

The first place of each leaderboard, with the date of its last update:

- Arena Code: WebDev: Claude Fable 5.2 (xHigh) (1,835.8; 2026-11-10)
- Artificial Analysis Coding Agent Index: Claude Code - Claude Opus 5.5 (Max, Default Fallback) (61.7; 2026-11-10)
- Scale SWE-Bench Pro V2 (hard): Opus 5 (Claude Code) xhigh (98; 2026-11-06)
- LiveBench 2.0: long-horizon: Claude Fable 5.2 (xhigh) (82.9; 2026-11-09)
- Arena Agent Arena: code: Claude Opus 5.5 (Max) (0.2; 2026-11-08)
- FrontierSWE V2: Claude Fable 5.2 (69.9; 2026-11-09)
- SWE-Marathon v1.1: Antigravity CLI - Gemini 4 Carbon (high) (53.8; 2026-11-07)
- LiveBench v1: agentic coding: deepseek-v4.1-flash-max (77.3; 2026-10-16)
- Design Arena: code overall: GPT-6 Astra (xhigh) (1,387; 2026-11-11)
- Design Arena: agentic web app, full-stack: Claude Opus 5.5 (1,408; 2026-11-11)
- Design Arena: agentic web app, frontend: GPT-6 Astra (max) (1,343; 2026-11-11)
- Design Arena: website: Muse Wonder (max) (1,367; 2026-11-11)
- Arena Code: Image-to-WebDev: Claude Opus 5.5 (Max) (1,751.9; 2026-11-10)
- WeirdML v3 (agentic): Codex - gpt-6-astra (xhigh) (42.2; 2026-11-11)
- GSO (software optimization): Claude Fable 5.1 (xhigh) (87.3; 2026-09-27)
- Scale SWE Atlas: codebase Q&A: Opus 5 (Claude Code) xHigh (63.2; 2026-09-09)
- Scale SWE Atlas: test writing: Fable-5.1 (Claude Code) xHigh (67; 2026-09-09)
- Scale SWE Atlas: refactoring: GPT 6 Astra (Codex) xHigh (59.1; 2026-09-09)
- Snorkel Agentic Coding 2.0: GPT-6 Astra (47.6; 2026-10-23)
- Snorkel SWE-bench CLI: Opus 5.5 (16.6; 2026-10-23)
- Vals AI: ProgramBench: Claude Fable 5.2 (high) (19; 2026-11-09)
- Vals AI: SRE-bench: GPT-6 Astra (max) (56.9; 2026-11-05)
- CWE-bench v2: Claude Opus 5.5 (Claude Code) (67.1; 2026-11-02)
- Endor Labs Agent Security League: Claude Code - Claude Fable 5.1 (36.9; 2026-10-24)
- Terminal-Bench-Science (Artificial Analysis run): Claude Fable 5.2 (Max, Default Fallback) (73.3; 2026-11-09)
- Design Arena: UI components: GPT-6 Astra (xhigh) (1,399; 2026-11-11)
- Design Arena: game dev: GPT-6 Astra (max) (1,466; 2026-11-11)
- Design Arena: mobile apps (agents): Claude Fable 5.2 (1,338; 2026-11-11)
- Design Arena: SVG: Arrow 2 Telos (1,500; 2026-11-11)
- Design Arena: data visualization: Claude Opus 5.5 (1,390; 2026-11-11)
- llm-stats: Best AI for coding: Claude Fable 5.2 (50.7; 2026-11-11)
- SWE-rebench: Fable 5 [high] (64.5; 2026-07-24)
- KiloBench (Terminal-Bench 2.0 in Kilo): OpenAI: GPT-6 Astra ($$$$) (79.3; 2026-09-28)
- Terminal-Bench-Science 0.1: GPT-6 Astra (Codex, max) (68.1; 2026-09-24)
- Snorkel LibraryDesignBench: Opus 5.5 (mini-SWE) (48.9; 2026-10-01)

Sources: [Arena Code: WebDev leaderboard](https://arena.ai/leaderboard/code/webdev) · [Artificial Analysis Coding Agent Index](https://artificialanalysis.ai/agents/coding-agents) · [SWE-Bench Pro (public) V2](https://labs.scale.com/leaderboard/swe_bench_pro_public_v2) · [LiveBench](https://livebench.ai/) · [Arena Agent Arena: Code](https://arena.ai/leaderboard/agent/code) · [FrontierSWE V2](https://www.frontierswe.com/) · [SWE-Marathon v1.1](https://www.swe-marathon.org/) · [Design Arena: Code overall](https://www.designarena.ai/leaderboard/code) · [Design Arena: agentic web dev](https://www.designarena.ai/leaderboard?section=agentic-web-dev) · [Design Arena: Website](https://www.designarena.ai/leaderboard/website) · [Arena Code: Image-to-WebDev](https://arena.ai/leaderboard/code/image-to-webdev) · [WeirdML v3 (Epoch AI Benchmarking Hub)](https://epoch.ai/data/benchmark_data.zip) · [GSO (software optimization)](https://gso-bench.github.io/assets/leaderboard.json) · [SWE Atlas (codebase Q&A, test writing, refactoring)](https://labs.scale.com/leaderboard/sweatlas-qna) · [Snorkel Agentic Coding 2.0](https://snorkel.ai/leaderboard/agentic-coding-2/) · [Snorkel SWE-bench CLI](https://snorkel.ai/leaderboard/swe-bench-cli/) · [Vals AI: ProgramBench](https://www.vals.ai/benchmarks/programbench) · [Vals AI: SRE-bench](https://www.vals.ai/benchmarks/srebench) · [CWE-bench v2](https://cwe-bench.com/) · [Endor Labs Agent Security League](https://www.endorlabs.com/research/ai-code-security-benchmark) · [Artificial Analysis: Terminal-Bench-Science](https://artificialanalysis.ai/evaluations/terminal-bench-science) · [Design Arena: UI components](https://www.designarena.ai/leaderboard/ui-components) · [Design Arena: Game dev](https://www.designarena.ai/leaderboard/game-dev) · [Design Arena: Mobile apps (agents)](https://www.designarena.ai/leaderboard/mobileapps) · [Design Arena: SVG](https://www.designarena.ai/leaderboard/svgs) · [Design Arena: data visualization](https://www.designarena.ai/leaderboard) · [Best AI for coding](https://llm-stats.com/leaderboards/best-ai-for-coding) · [SWE-rebench](https://swe-rebench.com/) · [KiloBench](https://kilo.ai/kilobench) · [Terminal-Bench-Science 0.1 (Snorkel mirror of the official board)](https://snorkel.ai/leaderboard/terminal-bench-science/) · [Snorkel LibraryDesignBench 1.0.0](https://snorkel.ai/leaderboard/librarydesignbench/)

### Can I run a coding model locally?

[Link to this question](https://coding-agent.josedacosta.net/en#q-local-coding-models)

Yes, on a powerful enough machine; the best choices by hardware, with their AA index and Terminal-Bench 4.0 (TB4) scores:

- 24 GB GPU (RTX 4090 / 3090): Qwen3.8-27B Q4_K_M, 16.5 GB, 38–46 tok/s, AA 33.7 · TB4 5.6%
- 32 GB GPU (RTX 5090): Qwen3.8-27B Q6 / NVFP4, 22–25 GB, 23–75 tok/s · MTP 98–134, AA 33.7 · TB4 5.6%
- 128 GB unified memory (Strix Halo, DGX Spark, M5 Max): Qwen3.8-Flash-Next IQ4_XS, 93.7 GB, 11–37 tok/s · MTP 47–84, AA 39.8 · TB4 25.3%
- 256 GB unified memory (M5 Ultra, M3 Ultra): GLM-5.3-Flash MLX 4-bit, 204 GB, 26–41 tok/s, AA 41.8 · TB4 32.8%

Sources: [Qwen3.8-27B hardware tests](https://hardware-corner.net/qwen3-8-27b-hardware-tests/) · [Qwen3.8-27B quantizations benchmarked](https://quesma.com/blog/qwen38-27b-quantizations-benchmarked/) · [Artificial Analysis model leaderboard](https://artificialanalysis.ai/leaderboards/models) · [Mac Studio M5 Ultra review](https://www.macstories.net)


## Share Coding Agent

Found it useful? Share the link in one click, or post the carousel that sums it up: every share credits the tool and links to it.

Opens the network with the text and the link filled in: you only confirm. The text is also copied, for the networks that drop it (Facebook, LinkedIn on phones).

The essentials of the page in slides, generated in your language, your mode and your color theme: change them at the top of the page and the slides follow. Every slide carries the name and the address of the tool.

[Share Coding Agent](https://coding-agent.josedacosta.net/en#share)

## Support and feedback

It started with a question people keep asking me: “What should I use right now?” I test constantly myself, roughly every week — there is always something new, in open source and open weights, in paid models, in model updates and new models, for images as for code. I was tired of repeating the same advice to dozens of people, so I made this page: I keep it up to date, and you can come back to see what to use or not, what to do or avoid. Your reminder of the moment.

The recommendations combine benchmark results and analyses with official provider guidance for each use case. Model and reasoning-effort choices take measured performance, cost and estimated quota consumption into account. Each choice links to dated sources, distinguishing independent measurements from vendor-published results.

This is a personal project. I usually update it as things happen, but on holiday or when other projects keep me busy, it can fall a little behind. If you notice it is out of date, send me a quick reminder — by email, WhatsApp, LinkedIn or any other way — and I will refresh everything right away. All my contacts are on my portfolio, at the root of this website.

Like most of my projects, this one is free and made with love. I hope it helped you; in any case, it was made with heart.

## References

Every figure on this page comes from these pages. For each one: what we took from it, and when.

- [Codex developer commands](https://learn.chatgpt.com/docs/developer-commands) — OpenAI; accessed 2026-09-29
- [Codex code review](https://learn.chatgpt.com/docs/code-review) — OpenAI; accessed 2026-10-10
- [Claude Code review](https://code.claude.com/docs/en/code-review) — Anthropic; accessed 2026-09-29
- [Claude Code: run and verify your app](https://code.claude.com/docs/en/skills) — Anthropic; accessed 2026-09-29
- [Codex prompting](https://learn.chatgpt.com/docs/prompting) — OpenAI; accessed 2026-09-29
- [Agent Security](https://cursor.com/docs/agent/security) — Cursor; accessed 2026-09-29
- [Introducing Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5) — Anthropic, 2026-09-28; accessed 2026-09-29
- [Claude Sonnet 5.5 System Card](https://www.anthropic.com/claude-sonnet-5-5-system-card) — Anthropic, 2026-09-28; accessed 2026-09-29
- [Claude Sonnet 5.5 (Claude Platform docs)](https://platform.claude.com/docs/en/models/sonnet-5-5/overview) — Anthropic, 2026-09-28; accessed 2026-10-10
- [Introducing Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5) — Anthropic, 2026-10-07; accessed 2026-10-10
- [Claude Haiku 5.5 System Card](https://www.anthropic.com/claude-haiku-5-5-system-card) — Anthropic, 2026-10-07; accessed 2026-10-10
- [Claude Haiku 5.5 (Claude Platform docs)](https://platform.claude.com/docs/en/models/haiku-5-5/overview) — Anthropic, 2026-10-07; accessed 2026-10-10
- [Monthly API credits for Max and Team plans](https://support.claude.com/en/articles/17154008) — Anthropic, 2026-10-07; accessed 2026-10-10
- [Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) — Anthropic, 2026-09-22; accessed 2026-09-28
- [Claude Opus 5.5 System Card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf) — Anthropic, 2026-09-22; accessed 2026-09-28
- [Claude Fable 5.1 and Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1) — Anthropic, 2026-09-01; accessed 2026-09-28
- [Introducing Claude Fable 5.2 and Claude Mythos 5.2](https://www.anthropic.com/claude-fable-and-mythos-5-2) — Anthropic, 2026-11-09; accessed 2026-11-11
- [Claude Fable 5.2 and Claude Mythos 5.2 System Card](https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card) — Anthropic, 2026-11-09; accessed 2026-11-11
- [Claude Fable 5.2 model page](https://platform.claude.com/docs/en/models/fable-5-2/overview) — Anthropic, 2026-11-09; accessed 2026-11-11
- [Introducing Claude Sonnet 5](https://www.anthropic.com/news/claude-sonnet-5) — Anthropic, 2026-06-30; accessed 2026-09-28
- [Models overview](https://platform.claude.com/docs/en/models/overview) — Anthropic; accessed 2026-11-11
- [Pricing](https://platform.claude.com/docs/en/about-claude/pricing) — Anthropic; accessed 2026-11-11
- [Effort](https://platform.claude.com/docs/en/build-with-claude/effort) — Anthropic; accessed 2026-11-11
- [Optimizing for cost and intelligence](https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence) — Anthropic; accessed 2026-10-10
- [Claude pricing](https://claude.com/pricing) — Anthropic; accessed 2026-11-11
- [How do usage and length limits work?](https://support.claude.com/en/articles/11647753) — Anthropic; accessed 2026-10-10
- [What is the Max plan?](https://support.claude.com/en/articles/11049741) — Anthropic; accessed 2026-11-11
- [What is the Ultra plan?](https://support.claude.com/en/articles/17728952) — Anthropic; accessed 2026-11-11
- [Claude Fable 5.2 launch offers](https://support.claude.com/en/articles/17506494) — Anthropic, 2026-11-09; accessed 2026-11-11
- [Claude Fable models on your plan](https://support.claude.com/en/articles/15424964) — Anthropic; accessed 2026-11-11
- [Models, usage, and limits in Claude Code](https://support.claude.com/en/articles/14552983) — Anthropic; accessed 2026-10-10
- [Model configuration](https://code.claude.com/docs/en/model-config) — Anthropic; accessed 2026-11-11
- [5-hour session limits increase](https://x.com/ClaudeDevs/status/2102438800836489554) — Anthropic (ClaudeDevs on X), 2026-09-22; accessed 2026-09-28
- [Claude Sonnet 5 System Card](https://www-cdn.anthropic.com/480e0bb54327b9622282e9c39a83a4f490ed377e/Claude%20Sonnet%205%20System%20Card.pdf) — Anthropic, 2026-06-30; accessed 2026-09-28
- [Introducing Claude Haiku 4.5](https://www.anthropic.com/news/claude-haiku-4-5) — Anthropic, 2025-10-15; accessed 2026-09-28
- [GPT-6 Astra](https://openai.com/index/gpt-6-astra/) — OpenAI, 2026-09-03; accessed 2026-09-28
- [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) — OpenAI, 2026-09-22; accessed 2026-10-10
- [Introducing GPT-6.1 Sol](https://openai.com/index/introducing-gpt-6-1-sol/) — OpenAI, 2026-09-29; accessed 2026-10-10
- [Introducing GPT-6.1 Terra](https://openai.com/index/introducing-gpt-6-1-terra/) — OpenAI, 2026-10-26; accessed 2026-10-27
- [GPT-6.1 Terra model page](https://developers.openai.com/api/docs/models/gpt-6.1-terra) — OpenAI, 2026-10-26; accessed 2026-10-27
- [GPT-6.1 Sol model page](https://developers.openai.com/api/docs/models/gpt-6.1-sol) — OpenAI, 2026-09-29; accessed 2026-10-10
- [DevDay 2026 Recap](https://openai.com/index/devday-2026-recap/) — OpenAI, 2026-09-29; accessed 2026-10-10
- [Codex speed (Fast and Ultrafast)](https://learn.chatgpt.com/docs/agent-configuration/speed) — OpenAI; accessed 2026-10-10
- [Codex rust-v0.161.0](https://github.com/openai/codex/releases/tag/rust-v0.161.0) — OpenAI, 2026-10-07; accessed 2026-10-10
- [API pricing](https://developers.openai.com/api/docs/pricing) — OpenAI; accessed 2026-10-27
- [GPT-6 Astra model page (and /gpt-6.1-sol, /gpt-6.1-terra, /gpt-6-sol, /gpt-6-luna)](https://developers.openai.com/api/docs/models/gpt-6-astra) — OpenAI; accessed 2026-10-10
- [Codex pricing](https://learn.chatgpt.com/docs/pricing) — OpenAI; accessed 2026-10-27
- [Codex models](https://learn.chatgpt.com/docs/models) — OpenAI; accessed 2026-10-27
- [Managing usage with GPT-6 Astra in Work and Codex](https://help.openai.com/en/articles/20001516-managing-usage-with-gpt-6-astra-in-work-and-codex) — OpenAI; accessed 2026-10-10
- [ChatGPT rate card (credit-based pricing)](https://help.openai.com/en/articles/11481834-chatgpt-rate-card-business-enterpriseedu-credit-based-pricing) — OpenAI; accessed 2026-10-10
- [About ChatGPT Pro tiers](https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers) — OpenAI, 2026-11-03; accessed 2026-11-11
- [API credits for ChatGPT subscribers](https://developers.openai.com/api/docs/guides/chatgpt-api-credits) — OpenAI, 2026-10-28; accessed 2026-11-11
- [Managing billing and seats in ChatGPT Business](https://help.openai.com/en/articles/8792536-managing-billing-and-seats-in-chatgpt-business) — OpenAI; accessed 2026-09-28
- [Weekly allowance value on Pro 20x (GPT-5.6 Sol vs GPT-6 Astra)](https://github.com/openai/codex/issues/43731) — GitHub user report, 2026-09-08; accessed 2026-09-28
- [Three weekly meters on Pro 5x](https://github.com/openai/codex/issues/43413) — GitHub user report, 2026-09-07; accessed 2026-09-28
- [Google AI subscriptions](https://gemini.google/subscriptions/) — Google; accessed 2026-11-11
- [Changes to Antigravity plans](https://antigravity.google/blog/changes-to-antigravity-plans) — Google, 2026-05-19; accessed 2026-10-10
- [Antigravity plans](https://antigravity.google/docs/plans/) — Google; accessed 2026-10-17
- [Antigravity models](https://antigravity.google/docs/models/) — Google; accessed 2026-11-11
- [Gemini API pricing](https://aistudio.google.com/docs/pricing) — Google, 2026-11-10; accessed 2026-11-11
- [Gemini thinking](https://ai.google.dev/gemini-api/docs/thinking) — Google; accessed 2026-11-11
- [Gemini 3.8 Flash evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_3-8_flash_model_evaluation.pdf) — Google DeepMind, 2026-09-02; accessed 2026-09-28
- [Gemini 4 Argon: our next era of frontier intelligence](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/) — Google, 2026-09-30; accessed 2026-10-10
- [Gemini 4 Argon: Built to solve your hardest problems](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/) — Google, 2026-10-13; accessed 2026-10-17
- [What's new in Gemini 4 Argon (Gemini API)](https://aistudio.google.com/docs/whats-new-gemini-4-argon) — Google, 2026-10-13; accessed 2026-11-11
- [Gemini 4 Argon in Google Antigravity](https://antigravity.google/blog/gemini-4-argon-in-google-antigravity) — Google, 2026-10-13; accessed 2026-10-17
- [Gemini 4 Argon model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Argon-Model-Card.pdf) — Google DeepMind, 2026-10-13; accessed 2026-10-17
- [Gemini 4 Argon model evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_4_argon_model_evaluation.pdf) — Google DeepMind, 2026-09-30; accessed 2026-10-10
- [Gemini 4 Carbon: our next frontier model](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/) — Google, 2026-10-28; accessed 2026-11-11
- [Gemini 4 Carbon is now available](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/) — Google, 2026-11-10; accessed 2026-11-11
- [What's new in Gemini 4 Carbon (Gemini API)](https://aistudio.google.com/docs/whats-new-gemini-4-carbon) — Google, 2026-11-10; accessed 2026-11-11
- [Gemini 4 Carbon and /swarm](https://antigravity.google/blog/gemini-4-carbon-in-google-antigravity) — Google, 2026-11-10; accessed 2026-11-11
- [Gemini 4 Carbon model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Carbon-Model-Card.pdf) — Google DeepMind, 2026-11-10; accessed 2026-11-11
- [Gemini 4 Carbon model evaluation](https://storage.googleapis.com/deepmind-media/gemini/gemini_4_carbon_model_evaluation.pdf) — Google DeepMind, 2026-11-10; accessed 2026-11-11
- [Fairwind Program](https://deepmind.google/fairwind-program/) — Google DeepMind; accessed 2026-10-17
- [Gemini API deprecations](https://aistudio.google.com/docs/deprecations) — Google, 2026-11-10; accessed 2026-11-11
- [Gemini 3.1 Pro model card](https://deepmind.google/models/model-cards/gemini-3-1-pro/) — Google DeepMind, 2026-02-19; accessed 2026-09-28
- [Grok 4.7](https://x.ai/news/grok-4-7) — xAI, 2026-09-21; accessed 2026-09-28
- [xAI API pricing](https://docs.x.ai/developers/pricing) — xAI; accessed 2026-09-28
- [Reasoning](https://docs.x.ai/developers/model-capabilities/text/reasoning) — xAI; accessed 2026-09-28
- [Grok plans](https://grok.com/plans) — xAI; accessed 2026-09-28
- [Grok FAQ](https://docs.x.ai/grok/faq) — xAI, 2026-08-27; accessed 2026-09-28
- [The X Subscription: pricing](https://docs.x.ai/x-subscription/pricing) — xAI; accessed 2026-10-17
- [Grok 4.7 model card](https://media.x.ai/v1/website/4p7card-5eccc980.pdf) — xAI, 2026-09-21; accessed 2026-09-28
- [Grok 4.6](https://x.ai/news/grok-4-6) — xAI, 2026-08-12; accessed 2026-09-28
- [Kimi K3](https://github.com/MoonshotAI/Kimi-K3) — Moonshot AI, 2026-07-16; accessed 2026-09-28
- [Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence](https://www.kimi.ai/blog/kimi-k3-1-pro) — Moonshot AI, 2026-10-24; accessed 2026-10-27
- [Kimi K3.1 Pro (Hugging Face)](https://huggingface.co/moonshotai/Kimi-K3.1-Pro) — Moonshot AI, 2026-11-02; accessed 2026-11-11
- [Kimi K3.1 Pro quickstart](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart) — Moonshot AI; accessed 2026-10-27
- [Kimi Code: What's New](https://www.kimi.com/code/docs/en/kimi-code/whats-new.html) — Moonshot AI, 2026-10-23; accessed 2026-10-27
- [Kimi API pricing](https://platform.kimi.ai/docs/pricing/chat) — Moonshot AI; accessed 2026-10-27
- [Kimi Code models](https://www.kimi.com/code/docs/en/kimi-code/models.html) — Moonshot AI; accessed 2026-10-27
- [Kimi Code membership](https://www.kimi.com/code/docs/en/kimi-code/membership.html) — Moonshot AI; accessed 2026-10-27
- [Kimi membership pricing](https://www.kimi.com/en/help/membership/membership-pricing) — Moonshot AI; accessed 2026-09-28
- [Coding plan tracker](https://codingplan.org) — codingplan.org, 2026-09-24; accessed 2026-09-28
- [Kimi plans in USD](https://geotoolbox.ai) — geotoolbox.ai, 2026-09-05; accessed 2026-09-28
- [GLM Coding Plan](https://z.ai/subscribe) — Z.ai; accessed 2026-09-28
- [GLM Coding Plan overview](https://docs.z.ai/devpack/overview) — Z.ai; accessed 2026-09-28
- [Z.ai API pricing](https://docs.z.ai/guides/overview/pricing) — Z.ai; accessed 2026-09-28
- [GLM-5.3 model guide](https://docs.z.ai/guides/llm/glm-5.3) — Z.ai; accessed 2026-09-28
- [GLM-5.3 launch post](https://z.ai/blog/glm-5.3) — Z.ai, 2026-08-14; accessed 2026-09-28
- [GLM-5.3 in Claude Code: real cost on the Coding Plan](https://www.decodeur-ia.com/articles/glm-5-3-claude-code-tuto-config-coding-plan-cout-reel-2026) — Décodeur IA, 2026-08-29; accessed 2026-09-28
- [GLM-5.3 model card](https://huggingface.co/zai-org/GLM-5.3) — Z.ai; accessed 2026-09-28
- [GLM-5.3-Flash](https://z.ai/blog/glm-5.3-flash) — Z.ai, 2026-08-26; accessed 2026-09-28
- [Claude Code usage meter measurements](https://github.com/anthropics/claude-code/issues/97074) — GitHub user reports, 2026-09-27; accessed 2026-09-28
- [Claude Max weekly limit measurements](https://dev.classmethod.jp/articles/claude-max-20x-weekly-limit-not-4x/) — Classmethod, 2026-08-20; accessed 2026-09-28
- [Cross-provider quota calibration](https://github.com/can1357/oh-my-pi/issues/13451) — oh-my-pi (GitHub), 2026-09-26; accessed 2026-09-28
- [Codex usage research](https://www.codexusage.dev/research) — codexusage.dev, 2026-09-22; accessed 2026-09-28
- [Pro 20x weekly meter measurement](https://github.com/openai/codex/issues/43222) — GitHub user report, 2026-09-10; accessed 2026-09-28
- [AI coding agent adoption 2026](https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/) — JetBrains Research, 2026-08-15; accessed 2026-09-28
- [GitHub Copilot plans](https://github.com/features/copilot/plans) — GitHub, 2026-09-28; accessed 2026-11-11
- [Models and pricing for GitHub Copilot](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing) — GitHub; accessed 2026-11-11
- [Gemini 4 Carbon in GitHub Copilot](https://github.blog/changelog/2026-11-10-gemini-4-carbon-in-github-copilot/) — GitHub, 2026-11-10; accessed 2026-11-11
- [Claude Fable 5.2 in GitHub Copilot](https://github.blog/changelog/2026-11-09-claude-fable-5-2-in-github-copilot/) — GitHub, 2026-11-09; accessed 2026-11-11
- [Route and orchestrate models with GitHub Copilot Fusion](https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion/) — GitHub, 2026-10-28; accessed 2026-11-11
- [Gemini 4 Argon in GitHub Copilot](https://github.blog/changelog/2026-10-13-gemini-4-argon-in-github-copilot/) — GitHub, 2026-10-13; accessed 2026-10-17
- [Model Studio Coding Plan](https://www.alibabacloud.com/help/en/model-studio/coding-plan) — Alibaba Cloud, 2026-09-28; accessed 2026-09-28
- [OpenCode Go](https://opencode.ai/docs/go/) — OpenCode, 2026-09-28; accessed 2026-11-11
- [MiniMax Token Plan](https://platform.minimax.io/docs/guides/pricing-token-plan) — MiniMax, 2026-09-28; accessed 2026-10-10
- [Muse Code](https://dev.meta.ai/products/muse-code) — Meta, 2026-09-28; accessed 2026-09-28
- [MiMo Token Plan](https://mimo.mi.com/docs/en-US/price/token-plan) — Xiaomi, 2026-09-21; accessed 2026-09-28
- [Mistral pricing](https://mistral.ai/pricing) — Mistral AI, 2026-09-28; accessed 2026-11-11
- [Cursor pricing](https://cursor.com/pricing) — Cursor, 2026-10-12; accessed 2026-10-17
- [What is The X Subscription?](https://cursor.com/help/the-x-subscription/overview) — Cursor, 2026-10-12; accessed 2026-10-17
- [Legacy Cursor plans](https://cursor.com/docs/account/pricing/legacy-cursor-plans) — Cursor; accessed 2026-10-17
- [Factory pricing](https://factory.com/pricing) — Factory, 2026-09-28; accessed 2026-09-28
- [Qwen3.8-27B hardware tests](https://hardware-corner.net/qwen3-8-27b-hardware-tests/) — Hardware Corner, 2026-08-17; accessed 2026-09-28
- [Qwen3.8-27B quantizations benchmarked](https://quesma.com/blog/qwen38-27b-quantizations-benchmarked/) — Quesma, 2026-08-26; accessed 2026-09-28
- [Mac Studio M5 Ultra review](https://www.macstories.net) — MacStories, 2026-09-21; accessed 2026-09-28
- [OpenRouter rankings: top models by task](https://openrouter.ai/rankings) — OpenRouter, 2026-09-27; accessed 2026-09-28
- [Anthropic Economic Index: impact on software development](https://www.anthropic.com/research/impact-software-development) — Anthropic, 2025-04-28; accessed 2026-09-28
- [How AI is transforming work at Anthropic](https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic) — Anthropic, 2025-12-02; accessed 2026-09-28
- [AIDev-pop: pull requests opened by coding agents](https://arxiv.org/html/2602.02345v1) — MSR 2026 (arXiv), 2026-02-03; accessed 2026-09-28
- [Artificial Analysis model leaderboard](https://artificialanalysis.ai/leaderboards/models) — Artificial Analysis; accessed 2026-11-11
- [Artificial Analysis evaluation details](https://artificialanalysis.ai/evaluations/terminalbench-4-0) — Artificial Analysis; accessed 2026-11-11
- [Announcing Artificial Analysis Intelligence Index v5](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5) — Artificial Analysis, 2026-10-27; accessed 2026-11-11
- [Artificial Analysis Coding Agent Index](https://artificialanalysis.ai/agents/coding-agents) — Artificial Analysis, 2026-11-10; accessed 2026-11-11
- [Arena Code: WebDev leaderboard](https://arena.ai/leaderboard/code/webdev) — Arena (LMArena), 2026-11-10; accessed 2026-11-11
- [Best AI for coding](https://llm-stats.com/leaderboards/best-ai-for-coding) — llm-stats (ZeroEval), 2026-11-11; accessed 2026-11-11
- [Terminal-Bench 4.1 leaderboard](https://www.tbench.ai/) — Terminal-Bench (tbench.ai), 2026-11-09; accessed 2026-11-11
- [Terminal-Bench 4.1](https://www.tbench.ai/news/terminal-bench-4-1) — Terminal-Bench (tbench.ai), 2026-11-09; accessed 2026-11-11
- [DeepSWE v1.1 leaderboard](https://deepswe.datacurve.ai/) — Datacurve; accessed 2026-09-28
- [CursorBench 4.0](https://cursor.com/cursorbench) — Cursor, 2026-11-10; accessed 2026-11-11
- [FrontierCode leaderboard data](https://cognition.com/data/frontiercode-leaderboard/data.json) — Cognition, 2026-11-09; accessed 2026-11-11
- [Vals AI benchmarks](https://www.vals.ai/benchmarks) — Vals AI, 2026-11-09; accessed 2026-11-11
- [Design Arena: agentic web dev](https://www.designarena.ai/leaderboard?section=agentic-web-dev) — Design Arena, 2026-11-11; accessed 2026-11-11
- [Claude Code best practices](https://code.claude.com/docs/en/best-practices) — Anthropic; accessed 2026-09-28
- [Claude Code changelog](https://code.claude.com/docs/en/changelog) — Anthropic, 2026-11-10; accessed 2026-11-11
- [Claude Code commands](https://code.claude.com/docs/en/commands) — Anthropic; accessed 2026-09-29
- [Claude Code workflows](https://code.claude.com/docs/en/workflows) — Anthropic; accessed 2026-09-28
- [Claude Code costs](https://code.claude.com/docs/en/costs) — Anthropic; accessed 2026-09-28
- [Claude Code permission modes](https://code.claude.com/docs/en/permission-modes) — Anthropic; accessed 2026-09-28
- [Claude Code in Chrome](https://code.claude.com/docs/en/chrome) — Anthropic; accessed 2026-09-28
- [Claude Code /goal](https://code.claude.com/docs/en/goal) — Anthropic; accessed 2026-09-28
- [Claude Code worktrees](https://code.claude.com/docs/en/worktrees) — Anthropic; accessed 2026-09-28
- [Prompting Claude Opus 5.5](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5) — Anthropic; accessed 2026-09-28
- [Codex best practices](https://learn.chatgpt.com/guides/best-practices) — OpenAI; accessed 2026-10-10
- [Antigravity CLI best practices](https://antigravity.google/docs/cli/best-practices) — Google; accessed 2026-09-28
- [Grok Build plan mode](https://docs.x.ai/build/features/plan-mode) — xAI; accessed 2026-09-28
- [Token Plan (Personal Edition)](https://www.alibabacloud.com/help/en/model-studio/token-plan-personal-overview) — Alibaba Cloud, 2026-09-28; accessed 2026-09-28
- [Token Plan (Team Edition)](https://docs.qwencloud.com/token-plan/team/token-plan-team-overview.md) — QwenCloud (Alibaba); accessed 2026-09-28
- [Qwen3.8-Max](https://www.qwencloud.com/models/qwen3.8-max) — QwenCloud (Alibaba); accessed 2026-09-28
- [Qwen3.8-Flash](https://www.qwencloud.com/models/qwen3.8-flash) — QwenCloud (Alibaba); accessed 2026-09-28
- [Deep thinking](https://docs.qwencloud.com/developer-guides/text-generation/thinking.md) — QwenCloud (Alibaba); accessed 2026-09-28
- [Qwen3.8-2.4T-A95B model card](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B) — Qwen (Alibaba), 2026-08-12; accessed 2026-09-28
- [Qwen3.8-Flash-Next model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) — Qwen (Alibaba), 2026-08-26; accessed 2026-09-28
- [Qwen Code](https://github.com/QwenLM/qwen-code) — Qwen (Alibaba), 2026-09-26; accessed 2026-09-28
- [Qwen Code settings](https://qwenlm.github.io/qwen-code-docs/en/users/configuration/settings/) — Qwen (Alibaba); accessed 2026-09-28
- [MiMo deep thinking](https://mimo.mi.com/static/docs/quick-start/usage-guide/text-generation/deep-thinking.md) — Xiaomi; accessed 2026-09-28
- [MiMo-V2.6-Pro model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL) — Xiaomi, 2026-09-21; accessed 2026-09-28
- [MiMo Code](https://github.com/XiaomiMiMo/MiMo-Code) — Xiaomi, 2026-09-26; accessed 2026-09-28
- [Design Arena: SVG](https://www.designarena.ai/leaderboard/svgs) — Design Arena, 2026-11-11; accessed 2026-11-11
- [Design Arena: data visualization](https://www.designarena.ai/leaderboard) — Design Arena, 2026-11-11; accessed 2026-11-11
- [MacBook SVG Benchmark: 34 AI Models, Ranked by the Laptop They Draw](https://playcode.io/blog/macbook-svg-benchmark) — Playcode, 2026-11-09; accessed 2026-11-11
- [SWE-Bench Pro (public) V2](https://labs.scale.com/leaderboard/swe_bench_pro_public_v2) — Scale AI, 2026-11-06; accessed 2026-11-11
- [SWE Atlas (codebase Q&A, test writing, refactoring)](https://labs.scale.com/leaderboard/sweatlas-qna) — Scale AI, 2026-09-09; accessed 2026-09-29
- [WeirdML v3 (Epoch AI Benchmarking Hub)](https://epoch.ai/data/benchmark_data.zip) — Epoch AI, 2026-11-11; accessed 2026-11-11
- [ALE-Bench (Epoch AI Benchmarking Hub)](https://epoch.ai/data/benchmark_data.zip) — Epoch AI, 2026-11-11; accessed 2026-11-11
- [Arena Agent Arena: Code](https://arena.ai/leaderboard/agent/code) — Arena, 2026-11-08; accessed 2026-11-11
- [Arena Code: Image-to-WebDev](https://arena.ai/leaderboard/code/image-to-webdev) — Arena, 2026-11-10; accessed 2026-11-11
- [Design Arena: Website](https://www.designarena.ai/leaderboard/website) — Design Arena, 2026-11-11; accessed 2026-11-11
- [Design Arena: Code overall](https://www.designarena.ai/leaderboard/code) — Design Arena, 2026-11-11; accessed 2026-11-11
- [Design Arena: UI components](https://www.designarena.ai/leaderboard/ui-components) — Design Arena, 2026-11-11; accessed 2026-11-11
- [Design Arena: Game dev](https://www.designarena.ai/leaderboard/game-dev) — Design Arena, 2026-11-11; accessed 2026-11-11
- [Design Arena: Mobile apps (agents)](https://www.designarena.ai/leaderboard/mobileapps) — Design Arena, 2026-11-11; accessed 2026-11-11
- [SWE-Marathon v1.1](https://www.swe-marathon.org/) — Abundant AI, 2026-11-07; accessed 2026-11-11
- [Snorkel Agentic Coding 2.0](https://snorkel.ai/leaderboard/agentic-coding-2/) — Snorkel AI, 2026-10-23; accessed 2026-11-11
- [Snorkel SWE-bench CLI](https://snorkel.ai/leaderboard/swe-bench-cli/) — Snorkel AI, 2026-10-23; accessed 2026-11-11
- [Snorkel LibraryDesignBench 1.0.0](https://snorkel.ai/leaderboard/librarydesignbench/) — Snorkel AI, 2026-10-01; accessed 2026-11-11
- [Terminal-Bench-Science 0.1 (Snorkel mirror of the official board)](https://snorkel.ai/leaderboard/terminal-bench-science/) — Snorkel AI / Laude Institute, 2026-09-24; accessed 2026-11-11
- [Artificial Analysis: Terminal-Bench-Science](https://artificialanalysis.ai/evaluations/terminal-bench-science) — Artificial Analysis; accessed 2026-11-11
- [Endor Labs Agent Security League](https://www.endorlabs.com/research/ai-code-security-benchmark) — Endor Labs, 2026-10-24; accessed 2026-11-11
- [CWE-bench v2](https://cwe-bench.com/) — Collinear AI, 2026-11-02; accessed 2026-11-11
- [GSO (software optimization)](https://gso-bench.github.io/assets/leaderboard.json) — GSO team (UC Berkeley), 2026-09-27; accessed 2026-10-11
- [Vals AI: ProgramBench](https://www.vals.ai/benchmarks/programbench) — Vals AI, 2026-11-09; accessed 2026-11-11
- [Vals AI: SRE-bench](https://www.vals.ai/benchmarks/srebench) — Vals AI, 2026-11-05; accessed 2026-11-11
- [FrontierSWE V2](https://www.frontierswe.com/) — Proximal Labs, 2026-11-09; accessed 2026-11-11
- [LiveBench](https://livebench.ai/) — LiveBench, 2026-11-09; accessed 2026-11-11
- [KiloBench](https://kilo.ai/kilobench) — Kilo, 2026-09-28; accessed 2026-09-28
- [SWE-rebench](https://swe-rebench.com/) — Nebius, 2026-07-24; accessed 2026-09-28
- [Vals Index](https://www.vals.ai/models/anthropic_claude-opus-5-5) — Vals AI; accessed 2026-11-11
- [Platforms and integrations](https://code.claude.com/docs/en/platforms) — Anthropic; accessed 2026-09-28
- [Use Claude Code Desktop](https://code.claude.com/docs/en/desktop) — Anthropic; accessed 2026-09-28
- [Use Claude Code in the cloud](https://code.claude.com/docs/en/claude-code-on-the-web) — Anthropic; accessed 2026-09-28
- [Claude Code for VS Code](https://marketplace.visualstudio.com/items?itemName=anthropic.claude-code) — Anthropic, 2026-09-25; accessed 2026-09-28
- [Codex glossary](https://learn.chatgpt.com/codex/glossary) — OpenAI; accessed 2026-09-28
- [Codex changelog: Codex joins the ChatGPT desktop app](https://learn.chatgpt.com/codex/changelog) — OpenAI, 2026-07-09; accessed 2026-09-28
- [Codex cloud](https://learn.chatgpt.com/codex/cloud) — OpenAI; accessed 2026-09-28
- [Codex IDE extension](https://learn.chatgpt.com/codex/ide) — OpenAI; accessed 2026-09-28
- [Codex CLI](https://learn.chatgpt.com/codex/cli) — OpenAI; accessed 2026-09-28
- [Google Antigravity docs: choose your surface](https://antigravity.google/docs/home) — Google; accessed 2026-09-28
- [Antigravity IDE Extensions](https://antigravity.google/docs/ide/extensions/) — Google; accessed 2026-09-28
- [Antigravity CLI installation](https://antigravity.google/docs/cli/install/) — Google; accessed 2026-09-28
- [Grok Build](https://docs.x.ai/build/overview) — xAI; accessed 2026-09-28
- [Grok Build headless and ACP](https://docs.x.ai/build/cli/headless-scripting) — xAI; accessed 2026-09-28
- [Kimi Code overview](https://www.kimi.com/code/docs/en/) — Moonshot AI; accessed 2026-09-28
- [Kimi Code Desktop changelog](https://www.kimi.com/code/docs/en/kimi-code-desktop/changelog.html) — Moonshot AI, 2026-09-23; accessed 2026-09-28
- [GLM Coding Plan supported tools](https://docs.z.ai/devpack/tool/others) — Z.ai; accessed 2026-09-28
- [Claude Code IDE Plugin](https://docs.z.ai/devpack/tool/claude-for-ide) — Z.ai; accessed 2026-09-28
- [Install ZCode](https://zcode.z.ai/en/docs/install) — Z.ai; accessed 2026-09-28
- [Qwen Code in VS Code](https://qwenlm.github.io/qwen-code-docs/en/users/integration-vscode/) — Qwen (Alibaba); accessed 2026-09-28
- [Qwen Code Desktop v0.24.6](https://github.com/QwenLM/qwen-code/releases/tag/desktop-v0.24.6) — Qwen (Alibaba), 2026-09-26; accessed 2026-09-28
- [About GitHub Copilot CLI](https://docs.github.com/en/copilot/concepts/agents/about-copilot-cli) — GitHub; accessed 2026-09-28
- [About GitHub Copilot cloud agent](https://docs.github.com/en/copilot/concepts/agents/coding-agent/about-coding-agent) — GitHub; accessed 2026-09-28
- [Installing the GitHub Copilot extension in your environment](https://docs.github.com/en/copilot/how-tos/set-up/install-copilot-extension) — GitHub; accessed 2026-09-28
- [Cursor CLI](https://cursor.com/docs/cli/overview) — Cursor; accessed 2026-09-28
- [Cloud Agents](https://cursor.com/docs/cloud-agent) — Cursor; accessed 2026-09-28
- [Cursor in JetBrains IDEs](https://cursor.com/docs/integrations/jetbrains) — Cursor; accessed 2026-09-28
- [Download OpenCode](https://opencode.ai/download) — Anomaly (OpenCode); accessed 2026-09-28
- [OpenCode web](https://opencode.ai/docs/web) — Anomaly (OpenCode); accessed 2026-09-28
- [Junie, the coding agent by JetBrains](https://www.jetbrains.com/junie/) — JetBrains; accessed 2026-09-28
- [npm downloads API: coding agent CLIs, week of 2026-10-04](https://api.npmjs.org/downloads/point/2026-10-04:2026-10-10/@openai/codex) — npm, 2026-10-10; accessed 2026-10-11
- [npm downloads API: coding agent SDKs, week of 2026-10-04](https://api.npmjs.org/downloads/point/2026-10-04:2026-10-10/@anthropic-ai/claude-agent-sdk) — npm, 2026-10-10; accessed 2026-10-11
- [VS Code Marketplace installs](https://marketplace.visualstudio.com/_apis/public/gallery/extensionquery) — Visual Studio Marketplace; accessed 2026-10-11
- [JetBrains Marketplace downloads](https://plugins.jetbrains.com/api/searchPlugins) — JetBrains Marketplace; accessed 2026-10-11
- [GitHub Action dependents](https://github.com/anthropics/claude-code-action/network/dependents) — GitHub; accessed 2026-10-11
- [Introducing Mistral Large 4](https://mistral.ai/news/mistral-large-4/) — Mistral AI, 2026-10-06; accessed 2026-10-10
- [Mistral Large 4 (Le Chonk): Built for Coders, Trusted by Defenders](https://mistral.ai/news/mistral-large-4-oss/) — Mistral AI, 2026-10-29; accessed 2026-11-11
- [Mistral Large 4 (Mistral docs)](https://docs.mistral.ai/models/mistral-large-4-0) — Mistral AI, 2026-11-05; accessed 2026-11-11
- [Vibe Code Web: limits and lifecycle](https://docs.mistral.ai/vibe/code/vibe-code-web/limits-and-lifecycle) — Mistral AI; accessed 2026-11-11
- [MiniMax M Plan](https://platform.minimax.io/docs/m-plan/intro) — MiniMax, 2026-09-30; accessed 2026-10-11
- [60 million Copilot code reviews and counting](https://github.blog/ai-and-ml/github-copilot/60-million-copilot-code-reviews-and-counting/) — GitHub, 2026-03-05; accessed 2026-09-29
- [About GitHub Copilot code review](https://docs.github.com/en/copilot/concepts/agents/code-review) — GitHub; accessed 2026-09-29
- [Asking GitHub Copilot questions in GitHub](https://docs.github.com/en/copilot/how-tos/chat-with-copilot/chat-in-github) — GitHub; accessed 2026-09-29
- [A Practical Approach to Verifying Code at Scale](https://alignment.openai.com/scaling-code-verification/) — OpenAI; accessed 2026-09-29
- [Review GitHub pull requests with Codex](https://developers.openai.com/codex/integrations/github) — OpenAI; accessed 2026-09-29
- [Claude Code GitHub Actions](https://code.claude.com/docs/en/github-actions) — Anthropic; accessed 2026-09-29
- [Bugbot](https://cursor.com/docs/bugbot) — Cursor; accessed 2026-09-29
- [Building a better Bugbot](https://cursor.com/blog/building-bugbot) — Cursor; accessed 2026-09-29
- [OpenCode GitHub](https://opencode.ai/docs/github) — Anomaly (OpenCode); accessed 2026-09-29
- [qwen-code-action](https://github.com/QwenLM/qwen-code-action) — Qwen (Alibaba); accessed 2026-09-29
- [Continue local sessions from any device with Remote Control](https://code.claude.com/docs/en/remote-control) — Anthropic; accessed 2026-09-29
- [Claude Code on mobile](https://code.claude.com/docs/en/mobile) — Anthropic; accessed 2026-09-29
- [Claude Code in Slack](https://code.claude.com/docs/en/slack) — Anthropic; accessed 2026-09-29
- [Run Claude Code programmatically](https://code.claude.com/docs/en/headless) — Anthropic; accessed 2026-09-29
- [Agent SDK overview](https://code.claude.com/docs/en/agent-sdk/overview) — Anthropic; accessed 2026-09-29
- [Asking GitHub Copilot questions in GitHub Mobile](https://docs.github.com/en/copilot/how-tos/chat-with-copilot/chat-in-mobile) — GitHub; accessed 2026-09-29
- [Copilot SDK](https://docs.github.com/en/copilot/how-tos/copilot-sdk) — GitHub; accessed 2026-09-29
- [Codex Remote](https://developers.openai.com/codex/remote) — OpenAI; accessed 2026-09-29
- [Use Codex in Slack](https://developers.openai.com/codex/integrations/slack) — OpenAI; accessed 2026-09-29
- [Use Codex in Linear](https://developers.openai.com/codex/integrations/linear) — OpenAI; accessed 2026-09-29
- [Codex SDK](https://developers.openai.com/codex/sdk) — OpenAI; accessed 2026-09-29
- [Non-interactive mode](https://developers.openai.com/codex/noninteractive) — OpenAI; accessed 2026-09-29
- [Cursor for iOS](https://cursor.com/docs/cloud-agent/mobile) — Cursor; accessed 2026-09-29
- [Cursor in Slack](https://cursor.com/docs/integrations/slack) — Cursor; accessed 2026-09-29
- [Cursor in Linear](https://cursor.com/docs/integrations/linear) — Cursor; accessed 2026-09-29
- [Using Headless CLI](https://cursor.com/docs/cli/headless) — Cursor; accessed 2026-09-29
- [Cursor SDK](https://cursor.com/docs/sdk/typescript) — Cursor; accessed 2026-09-29
- [OpenCode SDK](https://opencode.ai/docs/sdk) — Anomaly (OpenCode); accessed 2026-09-29
- [OpenCode CLI](https://opencode.ai/docs/cli) — Anomaly (OpenCode); accessed 2026-09-29
- [Antigravity Remote Control](https://antigravity.google/docs/remote-control/) — Google; accessed 2026-09-29
- [Qwen Code Channels](https://qwenlm.github.io/qwen-code-docs/en/blog/quickstart/channels-weixin-launch-announcement/) — Qwen (Alibaba); accessed 2026-09-29
- [Headless Mode](https://qwenlm.github.io/qwen-code-docs/en/users/features/headless/) — Qwen (Alibaba); accessed 2026-09-29
- [Qwen Code TypeScript SDK](https://qwenlm.github.io/qwen-code-docs/en/developers/sdk-typescript/) — Qwen (Alibaba); accessed 2026-09-29
- [Kimi Agent SDK](https://github.com/MoonshotAI/kimi-agent-sdk) — Moonshot AI; accessed 2026-09-29
- [ZCode Remote Control](https://zcode.z.ai/en/docs/remote-control) — Z.ai; accessed 2026-09-29
- [ZCode Bot Channel](https://zcode.z.ai/en/docs/bot-channel) — Z.ai; accessed 2026-09-29

- [Practical coding-agent workflows](https://coding-agent.josedacosta.net/en#workflows)
