Chess Puzzles
Games · 2025-12-08
Chess-reasoning benchmark with 100 novel engine-generated puzzles. Scores measure whether models identify the correct moves or solutions. Atlas keeps task versions 1.x here; Epoch's October 2026 2.0.x reruns are a separate benchmark, chess-puzzles-v2, because Epoch published no changelog and scores moved sharply.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.5 Pro | 64.0 |
| Gemini 3.1 Pro Preview | 55.0 |
| GPT-5.5 | 54.0 |
| Gemini 3 Flash Preview | 38.0 |
| Claude Opus 5 | 33.0 |
| Gemini 3 Pro Preview | 31.0 |
| Grok 4 | 28.0 |
| Qwen3.6 35B A3B | 26.0 |
| Gemini 3.1 Flash-Lite | 25.0 |
| Qwen3.7-Plus (source unspecified) | 24.0 |
| Qwen3.7 Flash | 23.0 |
| Qwen3.6 27B | 22.0 |
| Gemini 2.5 Pro | 20.0 |
| gpt-oss-120b | 20.0 |
| Kimi K2 Thinking | 20.0 |
| Qwen3.6-Max-Preview | 20.0 |
| Inkling-Small | 18.0 |
| Opus 4.6 | 17.0 |
| DeepSeek-V3.2 | 14.0 |
| Sonnet 4.6 | 13.0 |
| Qwen3.5 397B A17B | 13.0 |
| Seed OSS 36B Instruct | 13.0 |
| GPT-5.5 Instant | 12.0 |
| Kimi K2.5 | 12.0 |
| Qwen3 235B A22B Thinking 2507 | 12.0 |
| Qwen3.5-9B | 12.0 |
| GLM-5 | 10.0 |
| Qwen3.5 35B A3B | 10.0 |
| Qwen3 30B A3B Thinking 2507 | 8.0 |
| Opus 4.1 | 7.0 |
| Gemma 4 26B A4B IT | 6.0 |
| GPT-4 Turbo | 6.0 |
| Gemma 4 31B IT | 5.0 |
| Qwen3 32B | 5.0 |
| Qwen3 8B | 5.0 |
| QwQ-32B | 5.0 |
| GPT-4 0613 | 4.0 |
| gpt-oss-20b | 4.0 |
| Qwen3 14B | 4.0 |
| Qwen3-30B-A3B | 4.0 |