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.
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 CodeAnthropic
39%npm 12.5M/wk
Claude Code CLI
Claude Code Desktop
Claude Code for VS Code
Claude Code on the web
Remote Control, Claude app (iOS, Android)
Claude Code GitHub Actions, GitLab CI/CD
Claude Code in Slack
Claude Agent SDK, claude -p
GitHub CopilotGitHub
21%npm 1.6M/wk
GitHub Copilot CLI
GitHub Copilot extension
Copilot Chat (github.com/copilot)
GitHub Mobile
Copilot code review, Copilot cloud agent
Slack, Microsoft Teams, Jira, Linear, Azure Boards
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. Point at a green ✓ to see its figure.
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.
>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.
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).
Tied choices — pick the one your subscription gives you:
Recommended
Claude Fable 5.2·effort high
Agent
Claude Code
Subscription
Claude Max ($100) or Ultra ($200): Fable may use up to 50% of the weekly limit (on Pro, paid credits only)
Provider
Anthropic
≈ 2.9 agent-hours per 5-hour window · ≈ 14 per week on Claude Max 5x (central estimate). Simulate your case
Recommended
Claude Opus 5.5(default)·effort xhigh
Agent
Claude Code
Subscription
Claude Pro ($20), Max ($100) or Ultra ($200)
Provider
Anthropic
≈ 1.7 agent-hours per 5-hour window · ≈ 13 per week on Claude Pro (central estimate). Simulate your case
How long, then what
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.
Alternative
GPT-6 Astra·effort xhigh
Codex · ChatGPT Plus ($20) or Pro ($100 / $200 / $500)
Codex: 55.3 on the AA index at $3.66; max adds only +1.5 points for +70% cost.
Alternative
Gemini 4 Carbon·effort high(default)
Antigravity · Google AI Pro ($19.99) or Ultra ($99.99 / $199.99)
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.
How to work
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.
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.
For a large feature, ask Claude to interview you and write the result to SPEC.md; plans are also saved in ~/.claude/plans.
Edit the plan with Ctrl+G, then implement in a fresh session at /effort medium.
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).
Features, simple fixes, reviews, small refactors — the work you do all day. A bug you cannot explain has its own card: Debugging.
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 Opus 5.5(default)·effort medium(default)
Agent
Claude Code
Subscription
Claude Pro ($20), Max ($100) or Ultra ($200)
Provider
Anthropic
≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate). Simulate your case
How long, then what
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.
Alternative
GPT-6.1 Sol(default)·effort medium
Codex · ChatGPT Plus ($20) or Pro ($100 / $200 / $500)
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).
Alternative
Gemini 4 Carbon·effort medium
Antigravity · Google AI Pro ($19.99) or Ultra ($99.99 / $199.99)
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).
Alternative
Claude Fable 5.2·effort medium(default)
Claude Code · Claude Pro ($20), Max ($100) or Ultra ($200)
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).
Alternative
GLM-5.3·effort max(default)
Claude Code or ZCode · GLM Coding Plan ($18 / $80 / $168)
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).
How to work
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.
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.
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.
Before the pull request: /code-review looks for bugs in a fresh subagent; /simplify cleans up the diff but does not look for bugs.
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.
A bug you cannot explain yet: find the cause, reproduce it, fix it and prove the fix. Diagnosis needs more reasoning than writing code.
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 Opus 5.5(default)·effort high
Agent
Claude Code
Subscription
Claude Pro ($20), Max ($100) or Ultra ($200)
Provider
Anthropic
≈ 1.8 agent-hours per 5-hour window · ≈ 14 per week on Claude Pro (central estimate). Simulate your case
How long, then what
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.
Alternative
Claude Fable 5.2·effort xhigh
Claude Code · Claude Pro ($20), Max ($100) or Ultra ($200)
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).
Alternative
Gemini 4 Carbon·effort high(default)
Antigravity · Google AI Pro ($19.99) or Ultra ($99.99 / $199.99)
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.
Alternative
GPT-6.1 Sol(default)·effort xhigh
Codex · ChatGPT Plus ($20) or Pro ($100 / $200 / $500)
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.
How to work
Start with /effort high; add ultrathink to a single prompt to run just that turn at xhigh.
Give the full error, the stack trace and the exact command that fails, then ask for a reproduction (a failing test) before any fix.
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.
Done when the reproduction test passes and the rest of the suite still does; then /code-review on the diff.
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%).
Landing pages, components, dashboards — where humans judge the result.
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 Opus 5.5(default)·effort medium(default)
Agent
Claude Code
Subscription
Claude Pro ($20), Max ($100) or Ultra ($200)
Provider
Anthropic
≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate). Simulate your case
How long, then what
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).
Alternative
Claude Fable 5.2·effort xhigh
Claude Code · Claude Pro ($20), Max ($100) or Ultra ($200)
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.
Alternative
Claude Sonnet 5.5·effort xhigh
Claude Code · Claude Pro ($20), Max ($100) or Ultra ($200)
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.
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.
How to work
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.
Name the styles to avoid (for example “no cream background, no pill buttons”) rather than asking to “avoid a generic AI look”.
Keep the dev server running and iterate with small prompts.
Use plan mode only for multi-page features or a design system.
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.
Multi-hour refactors, migrations, terminal-heavy work where the agent runs on its own.
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.
Tied choices — pick the one your subscription gives you:
Recommended
Claude Opus 5.5(default)·effort xhigh
Agent
Claude Code
Subscription
Claude Max ($100) or Ultra ($200) for long sessions
Provider
Anthropic
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). Simulate your case
Recommended
GPT-6.1 Sol(default)·effort xhigh
Agent
Codex
Subscription
ChatGPT Plus ($20) or Pro
Provider
OpenAI
≈ 3.7 agent-hours per 5-hour window · ≈ 26.3 per week on ChatGPT Plus (Codex) (central estimate). Simulate your case
How long, then what
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.
Alternative
Claude Sonnet 5.5·effort max
Claude Code · Claude Pro ($20), Max ($100) or Ultra ($200)
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).
Alternative
GPT-6 Astra·effort high
Codex · ChatGPT Plus ($20) or Pro ($100 / $200 / $500)
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%.
Alternative
Gemini 4 Carbon·effort high(default)
Antigravity · Google AI Pro ($19.99) or Ultra ($99.99 / $199.99)
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.
Alternative
Claude Fable 5.2·effort high
Claude Code · Claude Pro ($20), Max ($100) or Ultra ($200)
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.
How to work
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.
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.
Effort: /effort xhigh in Claude Code; model_reasoning_effort = 'xhigh' for GPT-6.1 Sol in Codex (its default is low).
Split independent parts: /batch <instruction> or claude --worktree <name> in Claude Code; worktrees or cloud tasks in Codex.
ultracode (or “use a workflow”) orchestrates dozens of agents but uses much more quota: try it on one directory first.
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.
Many small, well-defined tasks: the best quality a small budget buys, without spending the quota of a frontier model on them.
Rename, reformat, add types across many files
Write docstrings, tests or translations in bulk
Subagents that search or summarize code
Quick one-line fixes
Tied choices — pick the one your subscription gives you:
Recommended
GPT-6.1 Terra·effort medium(default)
Agent
Codex
Subscription
ChatGPT Plus ($20)
Provider
OpenAI
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). Simulate your case
Recommended
Claude Haiku 5.5·effort medium(default)
Agent
Claude Code
Subscription
Claude Pro ($20), Max ($100) or Ultra ($200)
Provider
Anthropic
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). Simulate your case
How long, then what
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.
Alternative
Claude Opus 5.5(default)·effort low
Claude Code · Claude Pro ($20), Max ($100) or Ultra ($200)
Inside Claude Code, Anthropic recommends low for subagents: 87.4% on its SWE-bench Pro subset at $0.12 per solved task.
Alternative
MiMo-V2.6-Pro·effort high(default)
MiMo Code · MiMo Token Plan ($6 / $16 / $50 / $100)
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).
Alternative
GPT-6 Luna·effort max
Codex · ChatGPT Plus ($20) or Pro ($100 / $200 / $500)
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)).
How to work
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.
Fan out: /batch <instruction> or a claude -p loop in Claude Code; ask Codex to spawn subagents.
Test the prompt on 2–3 files before fanning out, and restrict tools with --allowedTools.
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.
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.
A very difficult algorithm, an optimization problem, a proof, a competitive-programming-style task.
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.
Tied choices — pick the one your subscription gives you:
Recommended
GPT-6 Astra·effort max
Agent
Codex
Subscription
ChatGPT Pro ($100 / $200 / $500)
Provider
OpenAI
≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate). Simulate your case
Recommended
GPT-6.1 Sol(default)·effort max
Agent
Codex
Subscription
ChatGPT Plus ($20) or Pro
Provider
OpenAI
≈ 3.7 agent-hours per 5-hour window · ≈ 26.3 per week on ChatGPT Plus (Codex) (central estimate). Simulate your case
How long, then what
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.
Alternative
Gemini 4 Argon·effort high(default)
Antigravity · Google AI Pro ($19.99) or Ultra ($99.99 / $199.99)
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.
Alternative
Claude Opus 5.5(default)·effort xhigh
Claude Code · Claude Pro ($20), Max ($100) or Ultra ($200)
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.
How to work
Ask first for a brute-force reference and tests, then for the optimized algorithm, and let the agent compare both.
In Codex: /model → GPT-6 Astra, model_reasoning_effort = 'max'; on ChatGPT Plus, GPT-6.1 Sol at max.
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.
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).
Simulations, numerical kernels, data science and research code where a wrong formula silently gives wrong results.
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.
Tied choices — pick the one your subscription gives you:
Recommended
Claude Fable 5.2·effort high
Agent
Claude Code
Subscription
Claude Max ($100) or Ultra ($200): Fable may use up to 50% of the weekly limit (on Pro, paid credits only)
Provider
Anthropic
≈ 2.9 agent-hours per 5-hour window · ≈ 14 per week on Claude Max 5x (central estimate). Simulate your case
Recommended
Claude Opus 5.5(default)·effort xhigh
Agent
Claude Code
Subscription
Claude Max ($100) or Ultra ($200) for long sessions
Provider
Anthropic
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). Simulate your case
How long, then what
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.
Alternative
Gemini 4 Carbon·effort high(default)
Antigravity · Google AI Pro ($19.99) or Ultra ($99.99 / $199.99)
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%).
Alternative
GPT-6.1 Sol(default)·effort max
Codex · ChatGPT Plus ($20) or Pro ($100 / $200 / $500)
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.
How to work
/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.
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.
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.
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.
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.
Tied choices — pick the one your subscription gives you:
Recommended
GPT-6 Astra·effort xhigh
Agent
Codex
Subscription
ChatGPT Plus ($20, 5–45 Astra messages per 5 hours) or Pro ($100 / $200 / $500)
Provider
OpenAI
≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate). Simulate your case
Recommended
Claude Opus 5.5(default)·effort high
Agent
Claude Code
Subscription
Claude Pro ($20), Max ($100) or Ultra ($200)
Provider
Anthropic
≈ 1.8 agent-hours per 5-hour window · ≈ 14 per week on Claude Pro (central estimate). Simulate your case
How long, then what
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.
Alternative
GPT-6 Sol·effort max
Codex · ChatGPT Plus ($20) or Pro ($100 / $200 / $500)
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.
Alternative
Gemini 4 Argon·effort high(default)
Antigravity · Google AI Pro ($19.99) or Ultra ($99.99 / $199.99)
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.
How to work
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.
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).
Data visualizations, dashboards and architecture or sequence diagrams written as code (ECharts, D3, Mermaid, PlantUML). Code generation only, not image generation.
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 Opus 5.5(default)·effort medium(default)
Agent
Claude Code
Subscription
Claude Pro ($20), Max ($100) or Ultra ($200)
Provider
Anthropic
≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate). Simulate your case
How long, then what
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.
Alternative
Claude Sonnet 5.5·effort xhigh
Claude Code · Claude Pro ($20), Max ($100) or Ultra ($200)
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.
Alternative
Kimi K3·effort high(default)
Kimi Code · Kimi Code plan (Plus or higher)
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).
Alternative
Gemini 4 Argon·effort high(default)
Antigravity · Google AI Pro ($19.99) or Ultra ($99.99 / $199.99)
On Google AI Pro ($19.99) in Antigravity: seventh on data visualization (1333). Effort not stated by the board; Antigravity defaults to High.
How to work
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.
Keep diagrams as text (Mermaid in Markdown) next to the code they describe, so the agent can update them with each change.
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.
Three.js scenes, WebGL, browser games, generative art and animation, written as code. Code generation only: this is not image or 3D-model generation.
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.
Tied choices — pick the one your subscription gives you:
Recommended
Claude Opus 5.5(default)·effort medium(default)
Agent
Claude Code
Subscription
Claude Pro ($20), Max ($100) or Ultra ($200)
Provider
Anthropic
≈ 2.3 agent-hours per 5-hour window · ≈ 17.6 per week on Claude Pro (central estimate). Simulate your case
Recommended
GPT-6 Astra·effort xhigh
Agent
Codex
Subscription
ChatGPT Pro ($100 / $200 / $500)
Provider
OpenAI
≈ 130.7 agent-hours per week on ChatGPT Pro 200 (Codex) (no 5-hour window; central estimate). Simulate your case
How long, then what
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.
Alternative
GPT-6 Astra·effort max
Codex · ChatGPT Plus ($20) or Pro ($100 / $200 / $500)
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.
Alternative
Claude Sonnet 5.5·effort xhigh
Claude Code · Claude Pro ($20), Max ($100) or Ultra ($200)
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).
How to work
Plan the scene structure first (plan mode), then build it step by step with the dev server running.
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.
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.
Whatever the task: what the Claude Code, Codex, Antigravity and Grok Build docs recommend, checked against their current versions (outdated tips left out).
01Verification first
Give the agent a pass/fail check (tests, build, screenshot) and ask for evidence, not claims. Every vendor puts this tip first.
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.
Worktrees, parallel sessions, subagents and workflows multiply the work done, but all draw on the same subscription limits: check the simulator before scaling out.
>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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
>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.
Models
Don’t: 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).
Don’t: 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.
Don’t: 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.
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.
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.
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.
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.
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.
Don’t: 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.
Don’t: 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.
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.
//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.
Score vs cost per task
API-equivalent cost at list prices, log scale — a secondary signal of relative consumption; subscription quota rules differ.
Scroll or pinch to zoom (Shift + scroll for the vertical axis), drag the slider, or use the zoom tool at the top right. Click a legend entry to hide or show it.
Loading the results…
Score by reasoning effort
Where the line flattens, the next effort level costs more quota for little or no gain.
//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.
Ordered by confidence (⭐⭐⭐ = independent, controlled, coding-focused and current; ⭐ = self-reported aggregates or stale sources), then by how widely the board is cited.
Arena Code: WebDev⭐⭐⭐
Blind pairwise human votes on web-development tasks (995,980 votes, cutoff 2026-11-10).
Confidence ⭐⭐⭐3 out of 3 stars of confidenceThe most cited coding arena: a very large number of blind human votes on real web apps, updated continuously.
Useful for: Front-end and web UI · 3D, games and creative coding as code
Source updated Nov 10, 2026 · Retrieved Nov 11, 2026 · 157 models
Anomaly detected: Claude Fable 5.2 (xHigh) enters first on only 994 votes (±22 points), and Gemini 4 Carbon (High) is fourth as a pre-release model. On 2026-11-10 Arena replaced the proprietary Mistral Large 4 entry with the open-weight model under the same name, so its two ratings are not comparable.
Loading the rows…
Show the data table
Artificial Analysis Coding Agent Index⭐⭐⭐
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.
Confidence ⭐⭐⭐3 out of 3 stars of confidenceIndependent, measures the agents themselves (Claude Code, Codex, Antigravity…) with their models, on current versions.
Useful for: Everyday coding · Long autonomous agent runs · High volume, subagents, quick edits
Source updated Nov 10, 2026 · Retrieved Nov 11, 2026 · 35 models
Incomplete coverage — this data does not include Claude Fable 5.2, Kimi K3.1 Pro, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: On 2026-10-27 Artificial Analysis replaced DeepSWE v1.1 with FrontierCode v1.1 in the index and re-scored SWE-Atlas-QnA: scores are not comparable with earlier ones. GPT-6.1 Terra scores the same at xhigh and max (48.2 and 48.1) for 56% of the cost at xhigh. Gemini 4 Argon's cost doubled on 2026-11-10 with an unchanged score, after AA doubled its API price. Muse Wonder Full and Muse Ember have no published cost: their cost per task is missing, not zero. Claude Fable 5.2 has no row yet.
Loading the rows…
Show the data table
Scale SWE-Bench Pro V2 (hard)⭐⭐⭐
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.
Confidence ⭐⭐⭐3 out of 3 stars of confidenceIndependent, real issues of commercial and open-source repositories re-graded on a clean image, current models.
Useful for: Debugging · Everyday coding
Source updated Nov 6, 2026 · Retrieved Nov 11, 2026 · 18 models
Incomplete coverage — this data does not include Claude Opus 5.5, Claude Fable 5.2, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Grok 4.7, Kimi K3.1 Pro, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Claude Sonnet 5.5 (mini-swe-agent, xhigh) scores 70.6% on the hard subset against 95.8% on the full set, a far larger drop than any other frontier model (Gemini 4 Argon 92.2 against 99.2). The hard subset publishes no confidence interval (0) except for Claude Haiku 5.5 (±13.0), and its 51 tasks move scores in steps of about 2 points: compare ranks, not small gaps. Inkling 2 Small (added 2026-11-06) is listed without a harness or an effort.
Loading the rows…
Show the data table
LiveBench 2.0: long-horizon⭐⭐⭐
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.
Confidence ⭐⭐⭐3 out of 3 stars of confidenceIndependent, contamination-limited long tasks, regularly refreshed, current models.
Useful for: Long autonomous agent runs · Everyday coding
Source updated Nov 9, 2026 · Retrieved Nov 11, 2026 · 33 models
Incomplete coverage — this data does not include Gemini 4 Carbon, GLM-5.3. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Each suite is a single run: between the 2026-10-25 and 2026-11-06 releases the same tasks moved some models by up to 22 points (Muse Spark 1.3 25.5 → 47.9), so read small gaps as noise. Claude Fable 5.2 is shown at xhigh because its max run scored lower overall; its cost per task ($160.93) is about twice Claude Opus 5.5's. LiveBench rewrote its earlier releases after publication (costs re-measured, models added). Mistral Large 4 appears twice: as "Mistral Large 4 1029" (24.7) and under its bare name (9.2).
Loading the rows…
Show the data table
Arena Agent Arena: code⭐⭐⭐
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.
Confidence ⭐⭐⭐3 out of 3 stars of confidenceReal coding-agent sessions rated by their users, close to a million sessions, current models.
Useful for: Everyday coding · Long autonomous agent runs
Source updated Nov 8, 2026 · Retrieved Nov 11, 2026 · 57 models
Incomplete coverage — this data does not include Claude Fable 5.2. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Arena changed two of the five signals behind this score: Steerability on 2026-10-01 (every judged turn instead of the turn after a correction) and Bash Recovery, replaced by Bash Reliability on 2026-10-12 (bash failures per call, detected by an LLM judge). Snapshots before 2026-10-16 are not comparable with this one. On 2026-11-08 four rows (DeepSeek-V4.1-Flash-20260910, DeepSeek-V4-Pro High, MiMo-V2.5-Pro and the proprietary Mistral Large 4) were dropped without explanation, and Mistral Large 4 now refers to the open-weight model.
Loading the rows…
Show the data table
FrontierSWE V2⭐⭐⭐
34 very hard engineering tasks with a 20-hour budget each — the long-horizon extreme.
Confidence ⭐⭐⭐3 out of 3 stars of confidenceIndependent, hard software-engineering tasks scored as the mean of five trials, current models.
Useful for: Long autonomous agent runs
Source updated Nov 9, 2026 · Retrieved Nov 11, 2026 · 24 models
Incomplete coverage — this data does not include GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Kimi K3.1 Pro, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: The source warns that Gemini 4 Argon's cost ($129.36 per trial) was measured with a much lower cache hit rate than in production (a non-production setting): its cost is overstated compared with real use.
Loading the rows…
Show the data table
SWE-Marathon v1.1⭐⭐⭐
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.
Confidence ⭐⭐⭐3 out of 3 stars of confidenceIndependent long tasks with many trials per model, recent and covering current models.
Useful for: Long autonomous agent runs
Source updated Nov 7, 2026 · Retrieved Nov 11, 2026 · 26 models
Incomplete coverage — this data does not include Claude Fable 5.2, GPT-6.1 Terra, Gemini 3.8 Flash, Kimi K3.1 Pro, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: On 2026-10-15 the site removed 11 older models from its board (Claude Opus 5, Opus 4.8, Sonnet 5, Fable 5, GPT-6 Sol, GPT-5.6 Sol/Terra/Luna, Grok 4.5, Grok 4.6, GLM 5.2); their results are unchanged at the source but, as on the site, they are no longer shown here. On 2026-10-20 it replaced Muse Spark 1.3's September runs with a new max run (46.25%, $14.29 per trial). Gemini 4 Carbon (Antigravity CLI, high, 53.75%), Muse Wonder Full and Muse Wonder lead with no published cost. Inkling 2 Small costs $424 per trial. Kimi K3 costs $661 per trial (133 M tokens), 4 to 10 times any other run: likely a pricing or token-counting problem at the source. Claude Haiku 5.5 at max (50.0%) is above Claude Opus 5.5 at max (46.25%), a gap of 6 trials out of 160 that is within noise.
Loading the rows…
Show the data table
LiveBench v1: agentic coding⭐⭐
LiveBench v1 (release 2026-06-25, now archived by LiveBench 2.0): contamination-limited questions; agentic coding in JavaScript, TypeScript and Python.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent and respected, but the v1 suite is now superseded by LiveBench 2.0.
Useful for: Everyday coding · High volume, subagents, quick edits
Source updated Oct 16, 2026 · Retrieved Nov 11, 2026 · 71 models
Incomplete coverage — this data does not include Claude Fable 5.2, GPT-6.1 Terra, Gemini 4 Carbon, Gemini 4 Argon, Kimi K3.1 Pro, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Design Arena: code overall⭐⭐
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.
Confidence ⭐⭐2 out of 3 stars of confidenceBlind human votes on generated code, independent, but fewer votes than Arena and judged on looks first.
Useful for: Front-end and web UI · SVG illustrations and icons as code · 3D, games and creative coding as code
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 195 models
Incomplete coverage — this data does not include Gemini 4 Carbon. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Design Arena: agentic web app, full-stack⭐⭐
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.
Confidence ⭐⭐2 out of 3 stars of confidenceBlind human votes on agent-built full-stack apps; independent, but a smaller and design-led vote.
Useful for: Front-end and web UI · Long autonomous agent runs
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 51 models
Incomplete coverage — this data does not include Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra, Gemini 4 Carbon, Gemini 4 Argon, Grok 4.7, GLM-5.3, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Design Arena has published no standard error since 2026-10-10 (empty for every row of all its boards), so no confidence interval is shown: compare ranks, not small Elo gaps. On 2026-11-11 Claude Opus 5.5 entered first on only 318 battles, and every other model lost about 19 Elo at once.
Loading the rows…
Show the data table
Design Arena: agentic web app, frontend⭐⭐
Same arena for multi-file React front-ends (13 tools).
Confidence ⭐⭐2 out of 3 stars of confidenceBlind human votes on agent-built front ends; independent, but a smaller and design-led vote.
Useful for: Front-end and web UI
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 56 models
Incomplete coverage — this data does not include Claude Fable 5.2, Claude Haiku 5.5, GPT-6.1 Terra, Gemini 4 Carbon, Gemini 4 Argon, Grok 4.7. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Design Arena has published no standard error since 2026-10-10, so no confidence interval is shown. GPT-6.1 Sol, fifth on 523 battles on 2026-10-27, fell to 14th (−39 Elo) on 871: early ranks on this board move a lot.
Loading the rows…
Show the data table
Design Arena: website⭐⭐
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.
Confidence ⭐⭐2 out of 3 stars of confidenceBlind human votes on websites; independent, but narrow and judged on looks.
Useful for: Front-end and web UI
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 206 models
Incomplete coverage — this data does not include Gemini 4 Carbon. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Arena Code: Image-to-WebDev⭐⭐
Blind pairwise votes on web pages built from an image, such as a mockup or a screenshot (210,832 votes, cutoff 2026-11-10).
Confidence ⭐⭐2 out of 3 stars of confidenceArena's blind votes, but narrow: turning a screenshot into a web page.
Useful for: Front-end and web UI
Source updated Nov 10, 2026 · Retrieved Nov 11, 2026 · 66 models
Incomplete coverage — this data does not include Claude Fable 5.2, Claude Haiku 5.5, GPT-6.1 Terra, Gemini 4 Carbon, Gemini 4 Argon, Kimi K3.1 Pro, GLM-5.3. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
WeirdML v3 (agentic)⭐⭐
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.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent and repeated runs, but machine-learning tasks rather than everyday software work.
Useful for: Algorithms, proofs and competitive programming · Scientific and numerical code
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 26 models
Incomplete coverage — this data does not include Claude Fable 5.2, Gemini 4 Carbon, GLM-5.3, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
GSO (software optimization)⭐⭐
Making real repositories faster without breaking them, in OpenHands, scored after removing reward hacks. Only GPT-6 Astra among current models.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent, hack-adjusted, but narrow (performance optimization) and updated less often.
Useful for: Algorithms, proofs and competitive programming
Incomplete coverage — this data does not include Claude Opus 5.5, Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Gemini 4 Argon, Gemini 3.8 Flash, Grok 4.7, Kimi K3.1 Pro, Kimi K3, GLM-5.3, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Rows dated 2026-09-27 use a new elicitation and jump to 70–88%: they are not comparable with older rows.
Loading the rows…
Show the data table
Scale SWE Atlas: codebase Q&A⭐⭐
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.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent (Scale), but a narrow skill (answering questions on a codebase) and not refreshed recently.
Useful for: Advanced planning and architecture · Everyday coding
Incomplete coverage — this data does not include Claude Opus 5.5, Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Gemini 4 Argon, Grok 4.7, Kimi K3.1 Pro, Kimi K3, GLM-5.3, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Scale SWE Atlas: test writing⭐⭐
Writing tests that catch real bugs in an existing codebase, in the vendor's agent.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent (Scale), but a narrow skill (writing tests) and not refreshed recently.
Useful for: Everyday coding · High volume, subagents, quick edits
Incomplete coverage — this data does not include Claude Opus 5.5, Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Gemini 4 Argon, Grok 4.7, Kimi K3.1 Pro, Kimi K3, GLM-5.3, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Scale SWE Atlas: refactoring⭐⭐
Refactoring real code without breaking its behavior, in the vendor's agent.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent (Scale), but a narrow skill (refactoring) and not refreshed recently.
Useful for: High volume, subagents, quick edits · Everyday coding
Incomplete coverage — this data does not include Claude Opus 5.5, Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Gemini 4 Argon, Grok 4.7, Kimi K3.1 Pro, Kimi K3, GLM-5.3, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Snorkel Agentic Coding 2.0⭐⭐
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.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent agentic coding tasks with costs, but no effort stated and fewer models.
Useful for: Long autonomous agent runs · Everyday coding
Source updated Oct 23, 2026 · Retrieved Nov 11, 2026 · 16 models
Incomplete coverage — this data does not include Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Kimi K3.1 Pro, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Snorkel SWE-bench CLI⭐⭐
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.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent multi-file repository changes, but no effort stated and fewer models.
Useful for: Debugging · Everyday coding
Source updated Oct 23, 2026 · Retrieved Nov 11, 2026 · 16 models
Incomplete coverage — this data does not include Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Kimi K3.1 Pro, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Vals AI: ProgramBench⭐⭐
Rebuilding real programs from scratch, run by Vals AI with one effort per model, with the cost per test.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent, building a program from scratch, but one effort per model and a narrow task.
Useful for: Long autonomous agent runs · Advanced planning and architecture
Source updated Nov 9, 2026 · Retrieved Nov 11, 2026 · 66 models
Incomplete coverage — this data does not include GPT-6.1 Terra, Gemini 4 Carbon, Kimi K3.1 Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Vals AI: SRE-bench⭐⭐
Reverse-engineering real, contamination-free binaries, run by Vals AI with one effort per model, with the cost per test.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent and contamination-free, but a narrow skill (reverse engineering) and one effort per model.
Useful for: Long autonomous agent runs
Source updated Nov 5, 2026 · Retrieved Nov 11, 2026 · 37 models
Incomplete coverage — this data does not include Claude Fable 5.2, GPT-6.1 Terra, Gemini 4 Carbon, Kimi K3, Qwen3.8 Max. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
CWE-bench v2⭐⭐
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.
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent security-fix tasks, recent, but narrow (vulnerabilities).
Useful for: Everyday coding
Source updated Nov 2, 2026 · Retrieved Nov 11, 2026 · 11 models
Incomplete coverage — this data does not include Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Gemini 3.8 Flash, Kimi K3, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: CWE-bench v2 (2026-11-02) replaced v1 with more tasks, more weakness types and stricter verifiers, so v1 scores do not compare. The top four (Claude Opus 5.5 67.1%, Muse Wonder 66.1%, Gemini 4 Argon 65.0%, GPT-6 Astra 64.2%) are within the ±5-point margin. Claude Fable 5.2 and Gemini 4 Carbon are not measured on v2.
Loading the rows…
Show the data table
Endor Labs Agent Security League⭐⭐
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).
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent, secure and correct code, but narrow (security) and fewer models.
Useful for: Everyday coding
Source updated Oct 24, 2026 · Retrieved Nov 11, 2026 · 34 models
Incomplete coverage — this data does not include Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Gemini 3.8 Flash, Grok 4.7, Kimi K3.1 Pro, Kimi K3, GLM-5.3, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Between 2026-10-16 and 2026-10-24 Endor Labs re-scored almost every row without a note, mostly downward (Claude Code + Claude Opus 5.5 33.5% → 30.7% secure), after adding anti-cheating safeguards to its method. Claude Opus 5.5 remains less often functionally correct than Claude Opus 5 (64.8% against 73.7%) and is now also less often secure (30.7% against 32.4%).
Artificial Analysis's own run of Terminal-Bench-Science with the same harness for every model, at several efforts (mean of 3 repeats).
Confidence ⭐⭐2 out of 3 stars of confidenceIndependent run with the same harness for every model, but scientific terminal tasks rather than software work.
Useful for: Scientific and numerical code
Source updated Nov 9, 2026 · Retrieved Nov 11, 2026 · 135 models
Anomaly detected: Claude rows were run with a default fallback model, and AA ranks Claude Opus 5.5 at xhigh (61.9%) above max (59.05%), while the Snorkel mirror gives Opus 5.5 max 63.3% in Claude Code. Claude Fable 5.2 scores higher at high (70.0%) than at xhigh (67.6%). GPT-6.1 Terra scores higher without reasoning (24.3%) than at low (7.6%).
Loading the rows…
Show the data table
Design Arena: UI components⭐⭐
Blind votes on interface components written as code (114,714 votes). No standard error is published, so close Elo scores cannot be told apart.
Confidence ⭐⭐2 out of 3 stars of confidenceBlind human votes on UI components; independent, but narrow and judged on looks.
Useful for: Front-end and web UI
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 186 models
Incomplete coverage — this data does not include Gemini 4 Carbon. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: The top two are 6 Elo apart: GPT-6 Astra (xhigh) on 669 battles and Kimi K3.1 Pro on 324. Claude Fable 5.2 entered 16th on only 125: treat these ranks as provisional.
Loading the rows…
Show the data table
Design Arena: game dev⭐⭐
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.
Confidence ⭐⭐2 out of 3 stars of confidenceBlind human votes on small games; independent, but narrow.
Useful for: 3D, games and creative coding as code
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 193 models
Incomplete coverage — this data does not include Gemini 4 Carbon. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Claude Sonnet 5.5, first on 435 battles on 2026-10-16, is now ninth on 2,008: early ranks on this board move a lot. Claude Fable 5.2 entered eighth on 396 battles.
Loading the rows…
Show the data table
Design Arena: mobile apps (agents)⭐⭐
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.
Confidence ⭐⭐2 out of 3 stars of confidenceBlind human votes on agent-built mobile apps; independent, but narrow and a smaller vote.
Useful for: Front-end and web UI
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 58 models
Incomplete coverage — this data does not include Claude Haiku 5.5, Gemini 4 Carbon, Gemini 4 Argon, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Claude Fable 5.2 enters first on only 179 battles, 40 Elo ahead of Claude Opus 5.5, and no standard error is published: treat its rank as provisional.
Loading the rows…
Show the data table
Design Arena: SVG⭐⭐
Blind votes on SVG icons, illustrations and logos written as code (211,612 votes). Includes SVG-specialized models that no subscription offers.
Confidence ⭐⭐2 out of 3 stars of confidenceBlind human votes on SVG drawings; independent, but a very narrow skill.
Useful for: SVG illustrations and icons as code
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 139 models
Incomplete coverage — this data does not include Claude Fable 5.2, GPT-6.1 Sol, GPT-6.1 Terra, Gemini 4 Carbon, Qwen3.8 Max. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Design Arena stopped publishing standard errors on 2026-10-10 (they were published for this board on 2026-09-29), so no confidence interval is shown: compare ranks, not small Elo gaps.
Loading the rows…
Show the data table
Design Arena: data visualization⭐⭐
Blind votes on charts and dashboards written as code (90,018 votes; no standard error published).
Confidence ⭐⭐2 out of 3 stars of confidenceBlind human votes on charts; independent, but a very narrow skill.
Useful for: Charts, dashboards and diagrams as code · Front-end and web UI
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 156 models
Incomplete coverage — this data does not include Claude Fable 5.2, Gemini 4 Carbon. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Since 2026-10-16 the board lists only models with at least 100 battles, a rule the site does not state. Claude Sonnet 5.5 is back second on 160 battles, and no standard error is published: treat its rank as provisional.
Loading the rows…
Show the data table
llm-stats: Best AI for coding⭐
Aggregate rating across 68 coding benchmarks, mostly self-reported by vendors.
Confidence ⭐1 out of 3 stars of confidenceAn aggregate that mixes many scores, most of them self-reported by the vendors: useful context, not a measurement.
Useful for: Everyday coding
Source updated Nov 11, 2026 · Retrieved Nov 11, 2026 · 207 models
Incomplete coverage — this data does not include Gemini 4 Carbon. Key current models are missing, so the ranking here is not fully relevant.
Anomaly detected: Muse Wonder Full and Muse Wonder are each listed twice on 2026-11-11, under two model ids with different ratings; the site gives no explanation. Claude Mythos 5.2 enters fourth on only two self-reported results, and Gemini 4 Carbon is in the index but hidden from the table.
Loading the rows…
Show the data table
SWE-rebench⭐stale
Fresh GitHub issues, decontaminated — but the last window ends on 2026-07-01, before most current models.
Confidence ⭐1 out of 3 stars of confidenceRespected method, but its last window predates most current models.
Incomplete coverage — this data does not include Claude Opus 5.5, Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Gemini 4 Argon, Gemini 3.8 Flash, Grok 4.7, Kimi K3.1 Pro, Kimi K3, GLM-5.3, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
KiloBench (Terminal-Bench 2.0 in Kilo)⭐
Terminal-Bench 2.0 run in the Kilo agent; cost is per full run of the suite, not per task.
Confidence ⭐1 out of 3 stars of confidenceRuns the older Terminal-Bench 2.0 in a single agent, and is not refreshed often.
Incomplete coverage — this data does not include Claude Opus 5.5, Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Gemini 4 Argon, Grok 4.7, Kimi K3.1 Pro, GLM-5.3, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Terminal-Bench-Science 0.1⭐
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.
Confidence ⭐1 out of 3 stars of confidenceVersion 0.1 mirror, superseded by the Artificial Analysis run of the same benchmark.
Useful for: Scientific and numerical code · Long autonomous agent runs
Incomplete coverage — this data does not include Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Gemini 4 Argon, Kimi K3.1 Pro, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
Loading the rows…
Show the data table
Snorkel LibraryDesignBench⭐
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.
Confidence ⭐1 out of 3 stars of confidenceNarrow (library API design), few models, not refreshed recently.
Useful for: Advanced planning and architecture
Source updated Oct 1, 2026 · Retrieved Nov 11, 2026 · 11 models
Incomplete coverage — this data does not include Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Carbon, Gemini 4 Argon, Gemini 3.8 Flash, Grok 4.7, Kimi K3.1 Pro, Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
#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.
Incomplete coverage — this data does not include Qwen3.8 Max, MiMo-V2.6-Pro. Key current models are missing, so the ranking here is not fully relevant.
>When running a model on your own machine makes sense: hardware, performance and limits compared with remote agents.
Running models locally
Local models are a separate tier: they cost no quota, but the best of them reach 5–33% on Terminal-Bench 4.0 versus about 60% for frontier models. Pair a local model for routine work with a remote frontier model for hard tasks.
>Focused answers to choose a model, balance reasoning effort and manage subscription quotas.
What is an AI 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.
Which agent, model and effort for “Advanced planning and architecture”?
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.
Which agent, model and effort for “Everyday coding”?
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.
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.
Which agent, model and effort for “Front-end and web UI”?
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).
Which agent, model and effort for “Long autonomous agent 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).
Which agent, model and effort for “High volume, subagents, quick edits”?
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).
Which agent, model and effort for “Algorithms, proofs and competitive programming”?
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).
Which agent, model and effort for “Scientific and numerical code”?
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.
Which agent, model and effort for “SVG illustrations and icons as code”?
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.
Which agent, model and effort for “Charts, dashboards and diagrams as code”?
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.
Which agent, model and effort for “3D, games and creative coding as code”?
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.
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.
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.
How many hours of agent work does a subscription give?
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).
>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.
Share the link
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.
>A personal project, free and made with care: why this page exists, and how to reach me.
Why this page
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.
How the recommendations are ranked
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.
Out of date? Tell me
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.
Share your feedback
Share your own experience too, so this page gets better for everyone: send me your feedback, even critical — I love criticism, it is what makes things better.
Free, made with heart ❤️
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.
A word for the providers 🚜
If you are one of the providers listed in this tool, feel free to contact me to offer quota resets or credits: that is what I miss most. Each of us already holds 1 to 4 Max subscriptions. Nothing would make us happier than resets or credits. Get in touch: you will get a special mention here ❤️.
pass@1 over 20 multi-hour tasks × 8 trials, mean cost per trial, tokens and errors for every agent × model × effort (data in the site's JS bundle, newest trial 2026-11-07; the deploy of 2026-10-15 hides 11 older models from the board, and so do we; Gemini 4 Carbon (Antigravity CLI) and Inkling 2 Small added in November 2026).
Coding Plan Pro $50: 6,000 requests per 5 hours, 45,000 per week, 90,000 per month; its model list has no Qwen3.8 model (up to qwen3.7-plus); sold in limited quantity, Lite discontinued.
`opencode web` starts a server (127.0.0.1 by default, network access protected by OPENCODE_SERVER_PASSWORD) and opens the interface in a browser; no hosted web surface.
Verification first (prompt, `/goal`, Stop hook, subagent); skip the plan for one-sentence changes; prune CLAUDE.md; `/clear` after two failed corrections; `/batch` fan-out; adversarial review.
2.1.299 (2026-10-14): idle compaction, a 200K+ token conversation left idle about 54 minutes is compacted before its one-hour prompt cache expires; minimum set by CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, off with "idleCompaction": false. 2.1.302 (2026-10-16): Idle compaction row in /config. 2.1.304 (2026-10-19): warning on Pro and Max when resuming an inactive conversation would use a noticeable share of the 5-hour limit. 2.1.311 (2026-10-26): ultrathink runs its turn at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 in the interactive terminal. 2.1.312 (2026-10-27): usage messages say "5-hour limit" instead of "session limit". 2.1.316 (2026-10-31): CLAUDE_CODE_DISABLE_ULTRATHINK, CLAUDE_CODE_SUBAGENT_EFFORT (effort of subagents when nothing else sets one), ultrathink at xhigh in VS Code; Opus 4.8 fast mode deprecated, removed 2026-11-30. 2.1.317 (2026-11-03): subagents on another model use that model's saved or default effort, not the session's. 2.1.321 (2026-11-09): Claude Fable 5.2 ($8 / $40, $0.20 cache reads, 1M context) becomes the fable alias; saved Fable 5.1 picks move to it on Pro, Max and Ultra; Ultra named alongside Pro and Max. 2.1.322 (2026-11-10): Fable 5.2 fast mode on Vertex (global, us and eu regions).
Local /code-review checks correctness; /review alias since 2.1.223. /simplify applies cleanup; --fix and --comment are explicit modification/publication options, not default review steps.
Bundled /run and /verify launch and exercise the app. /run-skill-generator records launch instructions; /verify may save a project skill. Availability depends on enabled bundled skills and project overrides.
Coding scores at max effort unless stated: SWE-bench Pro 93.0%, Multilingual 96.6%, Multimodal 74.1%; DeepSWE v1.1 74.3%; Terminal-Bench 4.0 67.6% (Claude Code --bare, 660 trials, ±1.8; Opus 5.5 66.4% at xhigh, Fable 5.1 55.8%); FrontierCode 2.0 49.1% at xhigh and 45.7% at max (44.4% if 22 timed-out attempts count as zero); FrontierSWE v2 69.9%.
From 2026-11-09: Pro and Team standard seats get every Fable model in their plan until 2026-11-16 (up to 50% of the weekly limit) plus one Fable reset; Max, Ultra and premium seats get one to three Fable resets by subscription length (under 6 months, 6–12, over 12), valid until 2027-03-08; a reset restores Fable usage on the weekly limit and refills the whole 5-hour window, without raising limits.
Fable 5.2, 5.1 and 5 on all paid plans; they use limits faster and are capped at 50% of the weekly limit on Max, Ultra and Team Premium; usage credits only on Pro and Team Standard; Fable 5.2 needs Claude Code 2.1.321 or later.
Model id claude-haiku-5-5, 1M context, 128K output; $0.10 input, $0.01 cache read, $0.50 output per MTok up to 100K-token prompts ($0.50 / $0.05 / $2.50 above); efforts low to max, medium by default; adaptive thinking.
Coding capabilities at max effort: SWE-bench Pro 64.8, Multilingual 83.7, Multimodal 30.7, FrontierCode Main 46.4 (45.8 at xhigh), Terminal-Bench 4.0 39.2, FrontierSWE v2 43.8; FrontierCode per effort in output tokens (figure 8.3.A).
Score and cost per task at each effort level for Opus 5.5, Fable 5.1 and Opus 5 on Terminal-Bench 4.0, FrontierCode v1.1 and CursorBench 4.0; five-hour limit change of 2026-09-22.
Prices of Free, Pro ($20, $17 a month billed yearly), Max ($100, 5× Pro), Ultra ($200, 10× Pro per week, no 5-hour limit for verified accounts), Team and Enterprise plans and what each includes; Fable on usage credits on Pro, up to half the weekly limit on Max and Ultra; monthly API credits on Max ($100), Ultra ($200) and Team (up to $500); API price table with Fable 5.2 $8 / $0.20 / $40.
Model id claude-sonnet-5-5, 1M context, 128K output; $2 input, $0.10 cache read (cut from $0.20 on 2026-10-07), $10 output per MTok; efforts low to max, default high on the API and medium in Claude Code and the apps; thinking can no longer be disabled (between_tools instead).
Effort levels available per model and when to use each (xhigh for agentic coding over 30 minutes, low for subagents); Haiku 5.5 supports low to max, medium by default; Fable 5.2 supports low to max, high by default on the API (start at high, gains above it are uneven).
Score and cost per task at each effort level for Fable 5.2, Opus 5.5 and Fable 5.1 on FrontierCode 2.0 (also GPT-6 Astra and Gemini 4 Carbon; Fable 5.2 49.1% at xhigh for $14.19, 45.6% at max for $44.70), WANDR, BenchCAD 2.1, Humanity's Last Exam and Chartography; benchmark table (FrontierSWE v2 69.9%, Terminal-Bench-Science 0.1 74.6%); $8 / $40, fast mode $16 / $80; medium is the default in the Claude apps and Claude Code; Fable up to 50% of the weekly limit for Pro and Team standard seats during the launch week.
Score and cost at each effort level for Haiku 5.5, Haiku 4.5, Sonnet 5.5 and GPT-6 Luna on OSWorld 2.1, GDPval-AA v2.1, Humanity's Last Exam and Terminal-Bench 4.0 (Sonnet 5.5 at $0.10 and $0.20 cache reads); headline Terminal-Bench 4.0 39.2 and FrontierCode 1.1 Main 46.4 at max; prices by prompt length; Sonnet 5.5 cache reads cut to $0.10; monthly API credits for Max and Team.
Score and cost at each effort level for Sonnet 5.5, Opus 5.5 and Sonnet 5 on Terminal-Bench 4.0 ($ per attempt), FrontierCode v1.1 and CursorBench 4.0 ($ per task); headline SWE-bench Pro 81.3 and Terminal-Bench 4.0 70.6 at max.
Default model (Opus 5.5 on every plan since v2.1.280; no Fable model is a default); default effort per model (medium on Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5); fable alias → Fable 5.2 and saved Fable 5.1 picks moved to it on Pro, Max and Ultra (v2.1.321); haiku alias → Haiku 5.5 since v2.1.293; ultrathink raises one terminal turn to xhigh (v2.1.311+, Fable 5.2 included).
Model IDs, release dates, context windows, max output and status of every current Claude model: current lineup Fable 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5; Fable 5.1, Fable 5 and Haiku 4.5 legacy, still available.
Max 5x $100, Max 20x $200, Team $20 per Standard and $100 per Premium seat pooled up to $500 a month; Claude Platform only (API, Managed Agents, Agent SDK with an API key); not interactive Claude Code nor extra usage; plan limits unchanged.
Cost per solved task on a 478-problem SWE-bench Pro subset for Opus 5.5, Fable 5.1 and Sonnet 5 configurations; effort trade-offs for Opus 5.5; Haiku 5.5: GPQA Diamond 85% at about a twentieth of Opus 5.5's cost per question.
Max at $100: five times Pro usage every five hours and every week, a separate weekly limit for Fable; monthly billing only; $100 a month of API credits; Max 20x replaced by the Ultra plan.
The Max 20x plan is now called Ultra, existing subscribers moved automatically with no price change; $200 a month, monthly only; 10 times Pro usage every week; no five-hour limit for verified accounts (a temporary limit until verification); a weekly limit across all models and a separate weekly limit for Fable; $200 a month of API credits.
Net improvement over an average model (with CI), sessions and the five signal scores of each model × effort on live coding-agent sessions (snapshot of 2026-11-08, 909,397 sessions, 57 models; flight payload; Bash Reliability replaced Bash Recovery on 2026-10-12).
Rating, 95% confidence bounds and votes of every model building a web page from an image (vote cutoff 2026-11-10, 210,832 votes, 66 models; entries[] in the payload).
Rating, 95% confidence interval, votes and rank of every model on blind pairwise web-development votes (entries[] in the server-rendered payload; vote cutoff 2026-11-10, 995,980 votes, 157 models). The page now sits behind a Cloudflare challenge: load it in a headed browser.
Intelligence Index v5.0.1 and AA's own Terminal-Bench 4.0 run for every model × effort, with cost per task and output tokens per task (JSON embedded in the page; 734 models, 119 with a v5 score, newest release 2026-11-09: Claude Fable 5.2). Costs AA publishes as exactly 0 (no API price) are left empty.
FrontierCode 2.0 (132 tasks, from 2026-10-27; data/frontier-code-2.0/data.json): score, pass rate, cost, tokens and flagged rate for every model × effort. FrontierCode v1 and v1.1 archive (frontiercode-leaderboard/data.json): score, pass rate and cost, main and extended subsets. Changelog: Gemini 4 Carbon 2026-10-28; SWE-2 Flash and Muse Spark Open 2026-11-04; Kimi K3.1 Pro 2026-11-05; Claude Fable 5.2 and Inkling 2 Small 2026-11-09.
Deterministic pass@1 and pass@4 with 95% CI and mean cost per rollout of 11 models in their agents, all at high reasoning, on the combined 250 tasks and on the public (100) and held-out (150) sets (v2, results as of 2026-11-02; table data attributes and data-v2.js). Archives: v1 at /v1 (120 tasks, programmatic and judge-panel pass@1/pass@4, updated 2026-10-28 with Gemini 4 Carbon) and v0 at /v0 (100 tasks).
The X Subscription (2026-10-12): Plus $8, Premium $30, Super $100, Ultra $200; one weekly pool for Cursor models shared with Grok (4×, 15×, 50× Plus); third-party models at API list price from prepaid credit on Premium and up.
Score, cost per task, tokens and steps for every model × effort (server-rendered results table). Rows added without a changelog entry: Gemini 4 Argon and Muse Wonder (after 2026-10-10), DeepSeek V4.1 Flash 1017 (after 2026-10-17), Gemini 4 Carbon and Claude Fable 5.2 (November 2026); the changelog's latest entry is still 2026-10-07, so the published date is inferred.
Hobby, Start, Pro $20, Pro+ $60, Ultra $200 and Teams continue for existing customers with two monthly pools (Cursor Models, Other Models); comparable X Subscription tiers Premium, Super, Ultra.
One subscription for Grok, Cursor and X, launched on 2026-10-12: Grok and Cursor share one weekly included usage pool; third-party models in Cursor (Claude, GPT) are not included and are billed at API list price from the first use, paid from prepaid credit; switching is optional and permanent, and existing Cursor, SuperGrok and X Premium plans continue.
Elo, win rate and battles for the full-stack and frontend agentic web-app arenas (full-stack from /api/leaderboard, frontend from the embedded board). No standard error published since 2026-10-10 (btStdErr null).
Elo, battles, win rate and mean generation time of 195 models across every code category (category allcategories, 1,950,799 votes; no standard error published). Claude Opus 5.5 369 s, Claude Sonnet 5.5 390 s, Claude Fable 5.2 732 s, GPT-6.1 Sol 228 s, GPT-6.1 Terra 132 s, GPT-6 Astra xhigh 574 s, max 1,093 s, Gemini 4 Argon 205 s, Kimi K3.1 Pro 1,121 s.
Elo, battles and win rate of 156 models on blind human votes for chart and dashboard code (POST /api/leaderboard, category dataviz, 90,018 votes; no standard error published; models with fewer than 100 battles are no longer listed since 2026-10-16).
Elo, battles, win rate and mean generation time of 193 models on browser games (category gamedev, 318,370 votes; no standard error published). Claude Opus 5.5 483 s, Claude Sonnet 5.5 486 s, Claude Fable 5.2 919 s, GPT-6.1 Sol 228 s, GPT-6.1 Terra 123 s, GPT-6 Astra xhigh 603 s, max 1,133 s, Gemini 4 Argon 198 s, Kimi K3.1 Pro 1,076 s.
Elo, battles and win rate of 139 models on blind human votes for SVG generation (POST /api/leaderboard, category svg, 211,612 votes; standard error no longer published since 2026-10-10).
Elo, battles and win rate of 206 models on single-shot websites (POST /api/leaderboard, category website, 1,190,870 votes; no standard error published).
ALE-Bench Performance score (algorithm optimization) and cost per model × effort (ale_bench_external.csv): GPT-6 Astra max 2951, GPT-6.1 Sol max 2744, GPT-6.1 Terra max 2609 at $0.41, GPT-6 Sol max 2462, Claude Opus 5.5 high 2147 (the only Opus 5.5 effort measured).
Copilot code review passed 60 million reviews; usage grew 10x since its April 2025 launch and accounts for more than one in five code reviews on GitHub.
Copilot cloud agent works in the background from the agents panel on GitHub.com, issues or VS Code and opens pull requests; it can also be started from Microsoft Teams and Slack (public preview) and from Azure Boards, Jira or Linear.
Claude Fable 5.2 generally available on 2026-11-09 for Copilot Pro+, Max, Business and Enterprise on the same surfaces; billed at provider list price; requires data retention by Anthropic (policy off by default for organizations; zero data retention for eligible enterprises until the end of 2026).
Gemini 4 Argon generally available on 2026-10-13 for Copilot Pro+, Max, Business and Enterprise in VS Code, Visual Studio, Copilot CLI, cloud agent, Copilot app, JetBrains, Xcode and Eclipse; billed at provider list price.
Gemini 4 Carbon generally available on 2026-11-10 for Copilot Pro+, Max, Business and Enterprise in VS Code, Visual Studio, Copilot CLI, coding agent, Copilot app, github.com, GitHub Mobile, JetBrains, Xcode and Eclipse; billed at provider list price; on by default for organizations unless disabled.
Copilot Pro, Pro+, Max, Business and Enterprise prices and the AI Credits included with each (unchanged: $10 / $39 / $100 with $15 / $70 / $200); Gemini 4 Carbon, Gemini 4 Argon and Claude Fable 5, 5.1 and 5.2 on Pro+ and Max, not on Free or Pro.
Copilot Fusion generally available on 2026-10-28 on all Copilot plans (VS Code, Copilot app, CLI, JetBrains): Standard, High and Max efforts route and orchestrate enabled models; billed per model used at its standard rate, 30% off for a limited time on paid plans; Fusion Max scored 60 on the AA Coding Agent Index at about half the cost per task of Claude Code with Opus 5.5 Max.
Models per plan: Gemini 4 Carbon (Low/Medium/High, default High) on Ultra and Enterprise only; Gemini 4 Argon (same levels) on paid Google AI Pro (or a trial with eligible premium devices), Ultra and Enterprise; Claude Opus 5.5 and Sonnet 5.5 (thinking) on Pro (non-trial), Ultra and Enterprise; Gemini 3.8 Flash on every plan; Gemini 3.7 Flash, 3.6 Flash and 3.1 Pro on Enterprise only; Claude 4.6 models and GPT-OSS-120b to be removed on 2026-11-18 (was 2026-11-02). Two usage pools with five-hour and weekly remaining percentages.
Remote Control: a browser dashboard, installable on a phone with push notifications, that drives Antigravity 2.0 and CLI sessions running on your own machines; no hosted web agent.
Gemini 4 Argon rolls out in the Antigravity desktop app and CLI for Google AI Pro and Ultra subscribers and eligible Gemini Enterprise subscriptions, from 2026-10-13.
General availability on 2026-10-13: Gemini API and AI Studio, Gemini app and AI Mode for Google AI Pro and Ultra, Antigravity; introductory price $2 / $10 per MTok (cached input 95% off) until 2026-11-30, then $4 / $20; 1M unified input-output token limit.
Restricted rollout to cyber defenders (Fairwind Program) before paid API customers and Google AI Ultra; introductory price $2 input / $10 output per MTok, cached input 95% off, then $4 / $20; 1M-token output limit; DeepSWE v1.1 77.9%.
Gemini 4 Carbon for Google AI Ultra subscribers in the Antigravity app, CLI and IDE extensions from 2026-11-10, and in Gemini for business; /swarm multi-agent teams in preview ("consume quota much faster than a standard chat").
General availability on 2026-11-10: Gemini API paid tier (AI Studio, Android Studio), Gemini for business; Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode (not Pro). Argon stays the choice for most work; Carbon uses more tokens. #1 on CursorBench 4.0, WANDR and AutomationBench; /swarm preview in Antigravity; intro price ends 2026-11-30.
Restricted rollout to a subset of cyber defenders (Fairwind Program) from 2026-10-28; same introductory price as Argon, $2 / $10 per MTok, cached input 95% off, then $4 / $20. Results table (same values as the evaluation PDF): Terminal-Bench 4.0 (AA) 65.2, FrontierSWE 2 61.2, SWE-sweep 9.6; WebDev Arena #4 (1768, +89); cheating on Terminal-Bench 4.0 reduced to 0.0%.
gemini-3.7-flash deprecated on 2026-10-08 with requests routed to gemini-3.8-flash; gemini-3.5-flash routed to gemini-3.6-flash; gemini-2.5-pro routed to gemini-3.8-flash and gemini-2.5-flash to gemini-3.5-flash-lite from 2026-12-01 (announced 2026-11-02); gemini-4-carbon (2026-11-10) and gemini-4-argon (2026-10-13) with no shutdown date.
Prices per million tokens: Gemini 4 Carbon and Gemini 4 Argon both $2 input / $0.10 cached / $10 output through 2026-11-30, then $4 / $0.20 / $20 (Batch and Flex half, Priority 1.8×, no long-context tier, not on the free tier; Contributor tier for Argon only); Gemini 3.8 Flash (introductory until 2026-12-31) and 3.1 Pro Preview. The page moved from ai.google.dev (301) and no longer prints a date; published = the Carbon entry of the API changelog.
Thinking levels and defaults per Gemini model (Gemini 4 Carbon and Gemini 4 Argon: low, medium, high, default medium); Gemini 4 shares one 1,048,576-token window between input and output.
Prices of Google AI Plus, Pro, Ultra 5x and Ultra 20x (unchanged: $4.99, $19.99, $99.99, $199.99); Gemini 4 Argon included from AI Pro, Gemini 4 Carbon on Ultra only ("highest access" to both); Antigravity rate-limit wording per plan.
Model id gemini-4-argon (stable); thinking levels low, medium (default, recommended for complex code and agentic use) and high; 1,048,576-token unified context, 65,536 output per synchronous request and up to 1M with long decode; implicit caching only (95% off); not on the free tier. Since Gemini 4 Carbon (2026-11-10): "efficient" model with fewer thinking tokens per turn than Carbon at the same price, and usable on the Google AI Pro plan in AI Studio (Carbon needs Ultra). The guide moved from ai.google.dev/gemini-api/docs/latest-model, now a model overview.
Model id gemini-4-carbon (stable, released 2026-11-10, no shutdown date); thinking levels low, medium (default) and high; same 1,048,576-token unified window, 65,536 output per synchronous request and up to 1M with long decode, and same per-token price as Argon; spends more thinking tokens per turn than Argon; implicit caching only; AI Studio subscription access on AI Ultra only (Argon on AI Pro and Ultra).
Vetted Fairwind partners (governments, critical infrastructure, core platforms) get Gemini 4 Argon without cyber guardrails, standalone or in CodeMender; they had exclusive access from 2026-09-30 until general availability on 2026-10-13.
Sparse mixture-of-experts, 1M combined input-output tokens, knowledge cutoff March 2026; distributed in the Gemini app, Gemini API, AI Studio, Gemini for business, Antigravity and AI Mode. Its results table is the same image as the 2026-09-30 evaluation PDF (no new figure).
Benchmark table of Gemini 4 Argon at its highest thinking setting versus GPT-6 Astra, Claude Fable 5.1 and Opus 5.5 (DeepSWE v1.1 77.9, Terminal-Bench 4.0 57.4, FrontierSWE v2 55.0, Vibe Code Bench 91.9), with the harness and settings per benchmark.
Based on Gemini 4 Argon; 1M combined input-output tokens; knowledge cutoff March 2026; uses more thinking tokens than Argon; Ultra in consumer surfaces and Antigravity, paid API tiers, Fairwind for verified defenders. Its results image is byte-identical to the evaluation PDF's.
Benchmark table of Gemini 4 Carbon and Argon at their highest thinking setting versus GPT-6 Astra, Claude Opus 5.5 and Fable 5.1: AA Index v5.0 59, Vals Index v2.2 70.4, SWE-sweep 9.6, FrontierSWE 2 61.2, Terminal-Bench 4.0 (AA) 65.2 (Argon 57.1), Agents' Last Exam v1.1 77.4, OSWorld-2.1 offline 81.5; harness and source per benchmark.
Hack-adjusted and raw Opt@1 scores per model × effort in OpenHands (leaderboard.json, last_updated 2026-09-27; rows after the 2026-09-27 elicitation change are not comparable with older ones).
Share of professional developers using each AI coding tool at work, May–July 2026 (Developer Ecosystem Survey 2026, 15,000+ respondents, reweighted by region): Claude Code 39%, GitHub Copilot 21%, Codex 16%, Cursor 12%, JetBrains AI and/or Junie about 9%, OpenCode 7%, Antigravity 6%.
LiveBench 2.0 releases of 2026-11-06 (shown as updated 2026-11-09), 2026-10-25 and 2026-10-14 (data/<release>/table.csv, categories.json, cost.csv, breakdown.json): suite scores, weighted overall (×2 long agentic, ×1.5 coding and design from 2026-10-25) and cost per task, each model at its highest supported effort (Claude Fable 5.2 at xhigh). Earlier releases were republished with changed costs and added models. LiveBench v1 release 2026-06-25 (under /v1/): coding and agentic-coding category scores and cost per question.
TrueSkill coding rating (μ − 3σ) over 68 benchmark games for every model (initialIndexData in the payload; 237 models reviewed). The CDN can serve stale copies even with a cache-busting query: check dateModified of each fetch.
M Plan replaced Token Plan on 2026-09-30 at the same prices: Go $22, Explore $55, Build $132 a month (minimax.io/m-plan); Explore 3× and Build 7.5× Go usage; 5-hour and weekly windows, no sizes; text model M3.1 Flash Preview.
Public preview of Mistral Large 4 (1T parameters, 52B active, open weights by the end of October); coding scores from the text: DeepSWE v1.1 61.7%, SWE-Atlas QnA 59.4%, Terminal-Bench 4.0 28.3%, Coding Agent Index 49.8% (runs by Artificial Analysis, effort not stated). Charts are images with rounded labels.
Open weights of the improved Mistral Large 4 ("Le Chonk") on 2026-10-29, three reasoning modes (low replaces none, medium, high by default), $0.70 / $0.07 cached / $2.10 per MTok. From the embedded chart JSON: Terminal-Bench 4.0 41.4% (28% for the preview), Artificial Analysis Coding Agent Index 48 at high (8.5M tokens per task) and 35 at medium (5.8M), Cognition's FrontierCode 2.0 19.7% (the high-effort run in Cognition's data, https://cognition.com/data/frontier-code-2.0/data.json: low 9.5, medium 8.8, Mistral Vibe harness), AA Cyber Index 50.9%.
Ids mistral-large-4 / mistral-large-4-0, version 26.10, GA, release date 2026-11-05, 1.05T total and 52B active parameters, 1M context, open weights (Mistral Ecosystem Open Weight License); $0.70 input, $0.07 cached, $2.10 output per MTok (preview list price $1.36 / $0.14 / $4.18). Reasoning guide: reasoning_effort low, medium or high (high recommended); none, minimal, xhigh and max refused. Mistral Medium 3.5 deprecated 2026-10-31 (retirement 2027-04-30), replaced by Mistral Large 4.
Pro $14.99 with $25.50 of API credits a month (was $30 on 2026-09-28; seen with France preselected in the country selector), Education $5.99 with $15, Team $24.99 per user ($50 minimum); fair usage, PAYG past the limit at API rates.
Free 2 and paid 100 Vibe Code Web sessions a day, 24-hour sessions; Vibe Code Web and the CLI with a Mistral model use Mistral Medium 3.5 (unchanged on 2026-11-11).
Input, cached input, output and cache-write prices per million tokens for Kimi K3.1 Pro ($2 / $0.20 / $10), K3.1 Pro Highspeed ($4 / $0.20 / $20), Kimi K3 ($3 / $0.30 / $15) and K2.7 Code; 1,048,576-token context for the K3 series.
Models in Kimi Code (K3.1 Pro, K3.1 Pro Highspeed at 2× quota, K3, K3 256K, K2.8 Preview, K2.7 Code Highspeed at 3× quota), effort levels with their defaults, and plan access.
K3.1 Pro series live in Kimi Code on 2026-10-23: K3.1 Pro on Plus / Moderato and above (1M context from Pro / Allegretto), K3.1 Pro Highspeed on Pro / Allegretto and above; low / high / max with high as the default.
Open weights of Kimi K3.1 Pro (repository created 2026-11-02, Kimi K3.1 Pro License), MXFP4 weights with quantization-aware training, and the launch benchmark table (Terminal-Bench 4.0 in the Claude Code harness).
Model ids kimi-k3.1-pro and kimi-k3.1-pro-highspeed, 1M context, thinking always on, reasoning_effort low / high / max with max as the API default, access after a $1 top-up.
Benchmark table of Kimi K3.1 Pro at max effort in the Claude Code harness (Terminal-Bench 4.0 65.7, DeepSWE v1.1 75.5, SWE-Marathon v1.1 51.7, Terminal-Bench-Science 0.1 38.1) against Fable 5.1, GPT-6 Astra, Opus 5.5, DeepSeek-V4.1-Flash-1017 and GLM-5.3; Terminal-Bench 4.0 score and cost per task at high (52.8%, $4.16) and max (65.7%, $5.25); K3.1 Pro Highspeed 2-3× faster; weights by 2026-11-04.
Pro 100, Pro 200 and Pro 500 ($500, Ultrafast). Updated 2026-11-03: Pro 100 and Pro 500 include twice their previous Work and Codex allowance; the Pro 200 reduction announced on 2026-09-29 is canceled, every Pro 200 subscription has the higher allowance (the lower one doubled to match); Plus unchanged; no five-hour limit on Pro.
Since 2026-10-28, monthly API credits for eligible subscriptions: Pro 100 $100, Pro 200 $200, Pro 500 $500, Business Standard $20 and Premium $100 per paid seat (capped at $500 per workspace); unused credits expire each cycle; Codex usage through API keys is not covered and the credits do not raise ChatGPT or Codex limits.
Standard, batch, flex, fast and ultrafast prices per million tokens for GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol and Luna; Ultrafast is 6× Standard (Astra and 6.1 Sol only).
Credits per million tokens for each model (25× the API price; GPT-6.1 Sol 50 / 2.5 / 250); Fast 2× and Ultrafast 6× when paid with credits; 2.5× and 8× of included usage.
Goal / Context / Constraints / Done when; start GPT-6.1 Sol at the client's default effort, Luna High, Astra Light; `/plan`, interview and PLANS.md; update AGENTS.md after a repeated mistake; worktrees for parallel tasks (the .md copy is gone: read the HTML page).
Review in CLI and desktop, local changes or a base branch; reviewer reports findings without changing the working tree. GitLab merge requests in preview and automatic cloud reviews.
CLI syntax for /init, /plan, /diff, /mention, /compact, /resume, /fork, /mcp and /skills; codex review --uncommitted includes staged, unstaged and untracked files. Review target flags and a custom prompt are mutually exclusive.
GPT-6.1 Sol recommended (Plus to Enterprise, not Free/Go), effort Light to Ultra; starting presets: client default for 6.1 Sol, Luna High, Astra Light; GPT-6.1 Terra for writing, visual understanding and computer use (Codex CLI 0.159.0 or later, desktop app, Work on the web, API); surfaces per model.
Plan prices, estimated local messages per 5 hours on Plus and Standard Business (GPT-6.1 Sol 15–160, GPT-6.1 Terra 35–285), no 5-hour limit on Pro; credit rates per model (GPT-6.1 Terra 25 / 1.25 / 125); Fast 2.5× and Ultrafast 8× of included usage.
Task context, expected and actual behaviour, constraints and verification for coding-agent requests. Used to write editorial workflow templates, not universal commands.
Pro 500 plan at 25× the Plus allowance with Ultrafast; Ultrafast up to 8× faster token generation (300 tokens per second) in Codex, 6× in the API; GPT-6.1 Sol launch; Codex cloud, CLI and Code Review updates.
Terminal-Bench 4.0 score, cost, output tokens and latency at each effort for GPT-6 Astra, GPT-5.6 Sol and Fable 5.1; FrontierCode 1.1 Extended; output tokens per DeepSWE task.
Model id gpt-6.1-sol; 1,050,000 context, 128,000 output, cutoff 2026-04-30; efforts low to max (no none), default medium; $2 input, $0.10 cached, $2.50 cache write, $10 output; Fast 2×, Ultrafast 6×, batch and flex 50%.
Score and cost per task at each effort for GPT-6.1 Sol, GPT-6 Sol and GPT-6 Astra on DeepSWE v1.1 (6.1 Sol 75.22 at high), Terminal-Bench Science 0.1, OSWorld 2.0, AutomationBench, GDP.pdf and factual errors; prices $2 / $0.10 / $10.
Score and cost per task at each effort for GPT-6.1 Terra, GPT-6.1 Sol, GPT-6 Astra and others on Terminal-Bench 4.0 (Terra 50.3 at max, $1.84), Terminal-Bench-Science 0.1, Agents' Last Exam v1, OSWorld 2.1, MMMU-Pro and factual errors; prices $1 / $0.05 / $5; Plus, Pro, Enterprise and Edu in Work and Codex.
Five-hour and weekly windows, shared allowance across Codex and Work, effort advice (Astra Low can beat Sol High); no 5-hour limit on Pro 100, 200 and 500; Ultrafast on Pro 500 at 8× included usage.
Go ($10), Go Plus ($40) and Go Teams ($20 per member, $5 a month more usage than Go on every model) prices, requests per 5 hours, week and month and monthly dollar limit per model (open-weight models plus Grok 4.7, GPT-6 Luna, Haiku 5.5 and Gemini 4 Argon at half price until 2026-11-30); DeepSeek V4.1 Flash 43,330 / 86,670 / 45,500 per 5 h. 5-hour limit 20% and weekly 50% of the monthly limit.
Share of spend and tokens per task over 30 days (code: general implementation 10.1%, debugging 5.2%, file edits 4.3%, review 2.8%, front-end UI 2.0%; math 0.8%), from /api/frontend/v1/rankings/task-spend.
Quality order judged by eye over 249 runs (34 models, first answer kept per model × effort), updated 2026-11-09: 1st Claude Opus 5.5 (best drawing at max on the Batch API, 255,957 tokens, $2.56; streamed max returns an empty file; xhigh $1.72, high $0.40), 2nd Claude Fable 5.2 (best at medium $3.91, high $3.05; xhigh and max streamed drew nothing, $5.12 each), 3rd GPT-6 Astra (max $2.29, every effort assembles; xhigh $0.98), 4th Claude Sonnet 5.5, 5th GPT-6.1 Sol (xhigh $0.21), 6th GPT-6 Sol (max $0.39), 7th Claude Fable 5.1, 8th GPT-6.1 Terra, 9th Gemini 4 Argon, 10th Kimi K3.1 Pro.
Score and confidence interval per model, harness and effort on the three SWE Atlas boards (sweatlas-qna, sweatlas-tw, sweatlas-refactoring); newest entry 2026-09-09.
Resolved % of the full and hard subsets per model, harness and effort (update of 2026-09-22: 642 tasks, web tools disabled, re-graded on a clean image; Sonnet 5.5 added 2026-10-12, Gemini 4 Argon and Muse Wonder 1.0 on 2026-10-14, Muse Spark Open on 2026-10-16, MiniMax M3.1 Flash on 2026-11-02, Inkling 2 Small on 2026-11-06). The hard subset is shown.
pass@1 and pass@5 with cost per trial and per task for 16 models (benchmark-data/tb_ed4_leaderboard.json loaded by the page; data as_of 2026-10-20, page modified 2026-10-23; no effort stated).
Score (pass rate² × simplicity) with 95% CI, cost per downstream problem and library cost per model × harness; reference arms left out of the layer (the unranked production-library arm scores 46.6).
pass@1 and pass@5 with cost per trial and per task for 16 models on multi-file changes (benchmark-data/swe_ed2_leaderboard.json; data as_of 2026-10-20, page modified 2026-10-23; no effort stated).
Resolution rate with CI per model × agent × effort, overall and per science domain, and the total cost of each run (shown rounded); official board at tbench.ai/leaderboard/terminal-bench-science/0.1 (rows loaded client-side).
Same 66 tasks with hardened verifiers; 4.0 submissions regraded from saved artifacts (162 of 11,550 trials changed, 28 kept at their 4.0 score); every 4.1 score is within its 4.0 95% CI.
Accuracy with confidence interval, pass@2–5, total cost and tokens per agent × model × effort over 330 trials (Terminal-Bench 4.1, the 2026-11-09 regrade of the 4.0 submissions; the 4.0 page now redirects to the home page).
Accuracy, cost per test and latency for Terminal-Bench 4.0 (every model re-run), Vibe Code Bench, Vals Index v2.2, IOI (competitive programming, with standard error), Vibe Code Bench 1-100 and other coding benchmarks (benchmark_view_*.json files linked from each page).
Vals Index v2.2 score, cost per test and latency (Finance Agent v2.1, EMB v1.1, Code Migration v1.1, HLAB-Vals-1.1 and all-pass Tax scoring; scores not comparable with v2.1).
Total installs on 2026-10-11: GitHub Copilot Chat 78,727,453; GitHub Copilot 74,729,595; Claude Code for VS Code 27,485,072; Codex (openai.chatgpt) 15,360,093; OpenCode (sst-dev.opencode) 1,205,522; Kimi Code 602,491; Google Antigravity (Google.google-antigravity) 557,580; Qwen Code Companion 376,437.
CursorBench 4.0 score, cost per task, output tokens and steps at each effort for Grok 4.7 and for Fable 5.1, Opus 5, Sonnet 5, GPT-5.6 Sol/Terra/Luna and Gemini 3.8 Flash (JS dataset behind the chart).
Lite $6 / Standard $16 / Pro $50 / Max $100 for 4.1B / 11B / 38B / 82B credits per month; no 5-hour or weekly window; credits per token (MiMo-V2.6-Pro 2.5 cache hit / 300 input / 600 output; V2.6-Flash 2 / 100 / 200); off-peak ×0.8.