Loading
90 days of the enterprise, pro and team plans free, no card needed. See the plans
Best models
Every model in the live reference catalogue with tool calls, ranked by the coding index (Artificial Analysis, 0-100) with vendor list prices per million tokens. 12 qualify today.
Strongest first; price breaks ties. Value is index points per blended dollar. Click a model for its full specification, or compare two side by side.
| # | Model | Coding index | Input / 1M | Output / 1M | Context | Value | Compare |
|---|---|---|---|---|---|---|---|
| 1 |  Claude Fable 5.1Anthropic | 82 | $10.0 | $50.0 | 1M | 4 | |
| 2 | Claude Opus 5Anthropic | 78 | $5.00 | $25.0 | 1M | 8 | vs Claude Fable 5.1 |
| 3 |  GPT-5.6 SolOpenAI | 77 | $4.00 | $20.0 | 1.1M | 10 | vs Claude Fable 5.1 |
| 4 | GPT-6 AstraOpenAI | 77 | $10.0 | $50.0 | 1.1M | 4 | vs Claude Fable 5.1 |
| 5 |  Grok 4.6xAI | 77 | $2.00 | $6.00 | 500K | 26 | vs Claude Fable 5.1 |
| 6 | GPT-5.6 TerraOpenAI | 77 | $2.00 | $12.0 | 1.1M | 17 | vs Claude Fable 5.1 |
| 7 | Claude Fable 5Anthropic | 77 | $10.0 | $50.0 | 1M | 4 | vs Claude Fable 5.1 |
| 8 |  Gemini 3.8 FlashGoogle | 76 | $0.82 | $4.13 | 1.0M | 46 | vs Claude Fable 5.1 |
| 9 |  Qwen3.8 Max (0902)Qwen | 76 | $2.00 | $6.00 | 1M | 25 | vs Claude Fable 5.1 |
| 10 |  Kimi K3Moonshot AI | 76 | $0.24 | $11.9 | 1.0M | 24 | vs Claude Fable 5.1 |
| 11 |  Muse Spark 1.3Meta | 76 | $1.25 | $4.25 | 1.0M | 38 | vs Claude Fable 5.1 |
| 12 | GPT-5.5OpenAI | 75 | $5.00 | $30.0 | 1.1M | 7 | vs Claude Fable 5.1 |
Coding is the task where the gap between models is widest and where it costs the most to get wrong: a wrong edit costs a review cycle, not a reread. The coding index aggregates code generation and repair evaluations (LiveCodeBench, SciCode and similar) and is the closest public measure of whether a model's patch will compile and pass. Tool calls are required here - an editor assistant that cannot run a formatter or read a file is a chat window, not a coding model.
Price matters more than the ranking suggests, because an editor assistant is output-heavy: completions, rewrites and explanations are generated tokens, which cost several times the input rate. A model two places down the list at a quarter of the output price often wins on a team's bill while the top model keeps the hard refactors.
That is the split the router makes per request. Short completions, renames and docstrings go to the cheaper qualifying model; a multi-file refactor or a failing test gets the strongest one. The rule below says exactly that, and the trace shows which rule fired.
Declare the models as targets and let the rule send the hard requests to the strongest one and the rest to the best value; the trace names the rule that fired.
# rules.yaml - coding: the strongest qualifying model for the hard requests, the best value for the rest
rules:
- name: coding-hard
when: { domains: [coding], actions: [code_generation, code_review], min_complexity: 0.45 }
prefer: [claude-fable-5.1]
weight: 0.8
- name: coding-default
when: { max_complexity: 0.45 }
prefer: [gemini-3.8-flash]
weight: 0.6