ChessBench
Games · 2026-03-22
ChessBench measures whether a language model can solve 150 fixed Lichess tactics from FEN and an ASCII board. Models must return the exact expected lowercase UCI move line; the headline strict accuracy counts malformed, missing, and incorrect lines as failures and averages 30 puzzles from each of five tactical tracks.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.8 Flash | 73.3 |
| Gemini 4 Argon | 70.0 |
| GPT-6.1 Sol | 68.7 |
| Claude Opus 5.5 | 65.3 |
| GPT-6 Sol | 62.0 |
| Gemini 3 Flash Preview | 48.7 |
| Claude Fable 5.1 | 45.3 |
| Grok 4.6 | 43.3 |
| Claude Haiku 5.5 | 41.3 |
| Claude Sonnet 5.5 | 39.3 |
| GPT-6 Luna | 36.7 |
| Haiku 4.5 | 8.7 |