Atlas

Benchmarks

← All benchmarks

ChessBench — Fork, Historical Output Repair

Games · 2026-03-25

The 30-puzzle Fork track under ChessBench's historical protocol permitting one follow-up request after a missing or unparseable output; repaired exact UCI lines count as correct. Retained separately from single-request runs.

Top models (higher is better)

ModelScore
Gemini 3.7 Flash56.7
Gemini 3.5 Flash53.3
Gemini 3.6 Flash50.0
GPT-6 Astra46.7
Gemini 3.1 Pro Preview43.3
Grok 4.536.7
Grok 4.1 Fast33.3
Grok 4.2033.3
Grok Build 0.133.3
Gemini 3.1 Flash Image Preview30.0
Claude Fable 523.3
GPT-5.6 Terra23.3
Grok 4.323.3
Opus 4.820.0
Claude Opus 520.0
GLM-520.0
Qwen3.6 Plus (2026-04-02)20.0
Sonnet 516.7
Opus 4.613.3
Opus 4.713.3
GLM-5.213.3
Qwen3.6 Plus Preview13.3
Gemini 2.5 Pro10.0
GPT-5.6 Sol10.0
GLM-5.16.7
GPT-5.46.7
Sonnet 4.63.3
GPT-5.53.3
GPT-5.6 Luna3.3
Loading Atlas data…