Models
The best AI model for coding right now
Two independent rankings, one measured by tests and one by people's votes, and what each model costs per token and per task.
In October 2026, the best AI model for coding on independent tests is Claude Opus 5.5. It leads Vals AI's Terminal-Bench 4.0 run at 65.2% and people's votes on LMArena's WebDev leaderboard at 1809, and costs $4 per million input tokens and $20 per million output. Claude Sonnet 5.5 (64.1%, third in votes) is the best value among the leaders at $2 and $10, and Moonshot's Kimi K3.1 Pro (63.6%) is the cheapest per task, at $4.57 against Opus 5.5's $13.20. GPT-6 Astra is second in votes and the fastest of the top models, and Claude Fable 5.1, Anthropic's largest model, ranks below Opus 5.5 on both. Which model is best still depends on the work and on the coding agent running it.
Key takeaways
- Best overall: Claude Opus 5.5, first on both rankings.
- Best value: Claude Sonnet 5.5 or GPT-6.1 Sol, both $2 / $10 per million tokens; GPT-6.1 Sol's tasks cost $1.72 on Vals AI's run.
- Cheapest per task near the top: Kimi K3.1 Pro, $4.57 a task at 63.6%, with open weights promised.
- Fastest of the leaders: GPT-6 Astra, about 36 minutes per Terminal-Bench task.
| Model | Price per million tokens | Terminal-Bench 4.0 (Vals AI) | Cost per task | LMArena WebDev |
|---|---|---|---|---|
| Claude Opus 5.5 | $4 / $20 | 65.2% | $13.20 | #1 (1809) |
| Claude Sonnet 5.5 | $2 / $10 | 64.1% | $16.51 | #3 (1769) |
| Kimi K3.1 Pro | $2 / $10 | 63.6% | $4.57 | #8 (1695, preliminary) |
| GPT-6 Astra | $10 / $50 | 59.6% | $9.58 | #2 (1774) |
| Claude Fable 5.1 | $10 / $50 | 58.1% | $17.18 | #5 (1745) |
| GPT-6.1 Sol | $2 / $10 | 55.1% | $1.72 | #4 (1752) |
| Claude Haiku 5.5 | $0.10 / $0.50 | 36.4% | $3.12 | Not in top 20 |
Rankings: Vals AI's Terminal-Bench 4.0 run, updated October 26, 2026, which runs every model in the same harness and reports cost per task; and LMArena's WebDev leaderboard, from more than 950,000 votes on web builds, October 27. Prices from Anthropic, OpenAI and Moonshot.
What is the best AI model for coding?
On the two independent rankings that cover the current models:
Vals AI, Terminal-Bench 4.0 (agents completing real terminal tasks, scored on whether they finish):
- Claude Opus 5.5: 65.15%, $13.20 a task.
- Claude Sonnet 5.5: 64.14%, $16.51.
- Kimi K3.1 Pro: 63.64%, $4.57.
- GPT-6 Astra: 59.60%, $9.58.
- Claude Fable 5.1: 58.08%, $17.18.
Gemini 4 Argon (57.58%), GPT-6.1 Sol (55.05%, $1.72) and Claude Opus 5 (53.53%) follow.
LMArena WebDev (people comparing two models' web builds and voting for the better one):
- Claude Opus 5.5 at max effort: 1809.
- GPT-6 Astra at max effort: 1774.
- Claude Sonnet 5.5 at xhigh effort: 1769.
- GPT-6.1 Sol at max effort: 1752.
- Claude Fable 5.1 at max effort: 1745.
Kimi K3.1 Pro is eighth (1695, preliminary) and Gemini 4 Argon tenth (1680).
Each lab's own launch posts show its own model ahead, often on benchmarks it chose. Independent runs that use one harness for every model are the fairer comparison.
Which model is best for each job?
- Hard, open-ended work (a large refactor, an unfamiliar codebase, a bug nobody can find): Claude Opus 5.5, or Claude Fable 5.1 when Opus at high effort still falls short, as Anthropic's docs suggest. See Fable 5.1 vs Opus 5.5.
- Everyday coding: Claude Sonnet 5.5 or GPT-6.1 Sol, at $2 and $10. Sonnet scores higher on both rankings; GPT-6.1 Sol's tasks cost a fraction as much.
- Speed: GPT-6 Astra finished Vals AI's tasks in about 36 minutes on average, against an hour or more for the Claude models.
- Budget at scale: Kimi K3.1 Pro, near the top for a third of Opus 5.5's cost per task. See Kimi K3 vs Claude.
- The cheapest models: Claude Haiku 5.5 (36.4% on Vals AI's run, $0.10 and $0.50 per million tokens) and GPT-6 Luna (13.6%, the same prices) suit quick edits and high-volume agents, not hard tasks.
- Open weights: Kimi K3 now, and K3.1 Pro once Moonshot publishes the weights (promised by November 4), under Moonshot's license.
What do they cost?
Per million input and output tokens:
- Claude Opus 5.5: $4 and $20.
- Claude Sonnet 5.5: $2 and $10.
- Claude Fable 5.1: $10 and $50 (Claude Fable pricing).
- GPT-6 Astra: $10 and $50, more on prompts over 272,000 tokens.
- GPT-6.1 Sol: $2 and $10, more on prompts over 272,000 tokens.
- Kimi K3.1 Pro: $2 and $10.
- Claude Haiku 5.5 and GPT-6 Luna: $0.10 and $0.50 (Haiku costs five times that on prompts over 100,000 tokens).
Cost per task depends on how many tokens a model spends, not just its rates. On Vals AI's run, Sonnet 5.5's tasks cost more than Opus 5.5's despite half the token price, and GPT-6 Astra's cost less than Fable 5.1's at the same token price.
Does the coding agent matter as much as the model?
Often more. The same model scores differently in Claude Code, Codex, Cursor or a minimal harness, and effort settings move scores by several points. Vals AI uses one simple agent for every model so the models can be compared; in daily use you'll run a model inside a full agent. For the agents themselves, see the best AI coding agents, and for the two labs side by side, Claude vs ChatGPT for coding.
Which coding models can you use for free?
- Claude's free plan includes Sonnet, but not Claude Code.
- ChatGPT's free plan includes Codex.
- Poly (usepoly.co) runs Claude Code and Codex in a shared browser room. Its free tier includes Claude Sonnet 5.5 and Haiku 5.5, GPT-6 Luna, Grok Build and Muse Spark 1.3, and paid plans add Opus 5.5, Fable 5.1, GPT-6 Astra and the rest. Each person in a room picks their own model, so a team can try several on the same code. What is Poly?
Common questions
What is the best AI model for coding right now?
Claude Opus 5.5, on independent rankings in October 2026: it leads Vals AI's Terminal-Bench 4.0 run at 65.2% and LMArena's WebDev votes. GPT-6 Astra is second in votes, and Claude Sonnet 5.5 and Kimi K3.1 Pro are close behind on tests for less.
Is Claude or GPT better for coding?
Claude models lead both independent rankings checked here, with Opus 5.5 first. GPT-6 Astra is second in people's votes and faster per task, and GPT-6.1 Sol is the cheapest per task among strong models.
What is the cheapest good AI model for coding?
GPT-6.1 Sol cost $1.72 per Terminal-Bench task on Vals AI's run at 55.1%, and Kimi K3.1 Pro $4.57 at 63.6%. Both cost $2 and $10 per million tokens, the same as Claude Sonnet 5.5.
Is Claude Fable 5.1 the best model for coding?
Not on independent tests. Fable 5.1 is Anthropic's largest model, but Opus 5.5 and Sonnet 5.5 score higher on Vals AI's Terminal-Bench 4.0 run and LMArena's WebDev votes, for less money. Anthropic suggests Fable for demanding long-horizon work.