ChessBench Accuracy
Games · 2026-07-28
ChessBench (chessbench.ai) has language models play full chess games against reference opponents and scores the moves; it is unrelated to the chess-bench.com puzzle benchmark carried as `chessbench`. Accuracy scores move quality against Stockfish as a weighted mean of branching, converting, and resisting accuracy (weights 10, 9, 1). Each part is scaled so that the random-move baseline is 0 and perfect play is 1, so the metric has a known ceiling but no floor: play worse than random scores below zero, and observed values reach -1.2. min_score is therefore left undeclared, which keeps this out of the score-level capability battery.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 85.0 |
| Gemini 4 Argon | 81.8 |
| Gemini 3.8 Flash | 80.8 |
| Gemini 3.1 Pro Preview | 80.7 |
| GPT-6.1 Sol | 80.6 |
| GPT-5 | 80.4 |
| Muse Spark 1.1 | 79.9 |
| Claude Opus 5.5 | 79.4 |
| Gemini 3.5 Flash | 79.4 |
| o3 | 79.4 |
| Gemini 3.7 Flash | 78.9 |
| Gemini 3 Flash Preview | 78.6 |
| GPT-5.5 | 78.1 |
| Gemini 3.6 Flash | 77.7 |
| GPT-6 Sol | 77.4 |
| Claude Fable 5 | 77.1 |
| Muse Spark 1.3 | 76.3 |
| Muse Spark 1.2 | 76.2 |
| Claude Sonnet 5.5 | 75.8 |
| GPT-5.1 | 75.7 |
| GPT-5.4 | 75.3 |
| GPT-5.6 Luna | 75.0 |
| GPT-5.6 Sol | 74.8 |
| Opus 4.8 | 72.7 |
| GPT-5.6 Terra | 72.1 |
| Opus 4.7 | 71.9 |
| Claude Fable 5.1 | 71.5 |
| GPT-5.2 | 71.4 |
| GPT-5 Mini | 70.1 |
| Gemini 3.1 Flash-Lite Preview | 69.5 |
| GPT-5 Nano | 68.5 |
| Claude Opus 5 | 67.8 |
| Sonnet 5 | 67.6 |
| Grok 4.6 | 67.6 |
| Claude Haiku 5.5 | 66.9 |
| GPT-5.4 Nano | 66.8 |
| Gemini 3.1 Flash-Lite | 66.7 |
| Grok 4.20 | 66.3 |
| GPT-6 Luna | 66.1 |
| O4 Mini | 64.1 |