LLM Reference

The coding leaderboard · for developers

Best for coding

6 editor picks · 16 eligible models · The model we'd hand a stranger reviewing a PR.

Editorial pick plus benchmark and API pricing context.

See raw /best
EDITOR'S CHOICEResearched 7d ago

Claude Fable 5

Anthropic · 1m context
Excellent

Top SWE-bench in production — patches like a senior engineer.

Anthropic's new flagship: 80.3% SWE-bench Pro, 96% SWE-bench Verified on Vals.ai, and 85.0% OSWorld-Verified make it the best production coding pick for non-trivial engineering tasks.

Open model
The numbers
$/1M out
$50.00
$10.00 input
Context
1m
max window
Pros
  • +80.3% SWE-bench Pro
  • +96% SWE-bench Verified on Vals.ai
  • +1M context window
Cons
  • −$50 / 1M out — double Opus 4.8
  • −Overkill for boilerplate or short generation

Also worth picking

The runners-up

ranked by editorial pick orderEditorial tiersExcellentStrongSolid
#2
Claude Opus 4.8
Anthropic · 1m
$25.00 / 1M out
Prior holder that remains a strong value at $25 out, with 69.2% SWE-bench Pro and 88.6% SWE-bench Verified.
#3
Claude Sonnet 5
Anthropic · 1m
$10.00 / 1M out
The new everyday Sonnet: 85.2% SWE-bench Verified, 63.2% SWE-bench Pro, 80.4% Terminal-Bench 2.1, and 128K max output at durable $3/$15 — ahead of GPT-5.5 for agentic coding loops we run daily.
#4
GPT-5.5
OpenAI · 1.05m
$30.00 / 1M out
OpenAI's current flagship: SWE-bench Pro 58.6 and HumanEval 94.2 — the best non-Claude reviewer we run on a real PR.
#5
DeepSeek V4 Pro
DeepSeek · 1m
$0.87 / 1M out
Tops LiveCodeBench (93.5) and 80.6 SWE-bench at $0.87 out, open weights — near-frontier coding for a fraction of the price.
#6
Claude Sonnet 4.6
Anthropic · 1m
$15.00 / 1M out
Solid predecessor Sonnet — keep until Sonnet 5 accumulates more third-party harness rows.

Eligibility

16 models are eligible for this board

Eligibility means tagged with useCases: [coding]. Pins must come from this pool.

All picks