Atlas

Benchmarks

← All benchmarks

Chess Puzzles (task version 2.0)

Games · 2026-10-13

Epoch AI's Chess Puzzles reruns under task version 2.0.x, published October 13-14, 2026. Epoch replaced most models' 1.x runs with these and published no changelog; several scores moved sharply (e.g. Claude Opus 4.7 max from 7 to 41), so the protocol is treated as distinct from 1.x. It is its own capability family, separate from the 1.x family: 80 models have 2.0 results and only 7 also have 1.x results, so the two protocols enter the fit side by side rather than pooled.

Top models (higher is better)

ModelScore
GPT-6 Astra Pro76.0
GPT-6 Astra74.0
GPT-5.6 Sol Pro68.0
GPT-6.1 Sol68.0
Claude Opus 5.567.0
Gemini 3.7 Flash66.0
GPT-5.563.0
GPT-5.5 Pro62.0
GPT-5.2 Pro60.0
Claude Sonnet 5.559.0
Gemini 4 Argon59.0
GPT-5.4 Pro59.0
GPT-5.6 Sol59.0
Gemini 3.8 Flash58.0
GPT-6 Sol56.0
Claude Haiku 5.555.0
Opus 4.852.0
Gemini 3.1 Pro Preview51.0
Claude Fable 5.148.0
DeepSeek V4.1 Flash48.0
GPT-5.248.0
GPT-5.448.0
GPT-5 Pro48.0
Claude Fable 547.0
GPT-5.6 Terra47.0
DeepSeek V4.1 Flash 101746.0
DeepSeek V4 Pro 081346.0
Sonnet 545.0
Gemini 3.5 Flash45.0
GPT-5.4 Nano45.0
Claude Opus 543.0
GPT-5.6 Luna43.0
Gemini 3.6 Flash42.0
GPT-5.142.0
Opus 4.741.0
Grok 4.741.0
Gemini 3 Flash Preview39.0
GLM-5.337.0
Kimi K337.0
Muse Spark 1.337.0
Loading Atlas data…