ChessBench — Historical Output Repair
Games · 2026-03-25
Historical ChessBench protocol over the same 150 fixed Lichess puzzles. The runner at each result's original publication revision permits one follow-up request when the first output has no parsed move line; a successfully repaired line can count as correct. Retained separately from the later single-request protocol. All unique puzzles remain binary scored items; repair attempts do not increase item count.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.7 Flash | 70.7 |
| GPT-6 Astra | 66.0 |
| Gemini 3.6 Flash | 65.3 |
| Gemini 3.5 Flash | 61.3 |
| Grok 4.5 | 60.7 |
| Grok 4.1 Fast | 58.7 |
| Grok Build 0.1 | 58.7 |
| Gemini 3.1 Pro Preview | 55.3 |
| Grok 4.20 | 54.7 |
| Grok 4.3 | 48.7 |
| Claude Fable 5 | 46.0 |
| Gemini 3.1 Flash Image Preview | 46.0 |
| GPT-5.6 Sol | 42.7 |
| GPT-5.6 Terra | 40.7 |
| Claude Opus 5 | 32.7 |
| GPT-5.4 | 32.7 |
| GPT-5.5 | 29.3 |
| Qwen3.6 Plus (2026-04-02) | 28.0 |
| Opus 4.8 | 24.0 |
| GPT-5.6 Luna | 24.0 |
| Qwen3.6 Plus Preview | 22.0 |
| Opus 4.7 | 18.7 |
| Sonnet 5 | 18.7 |
| Opus 4.6 | 16.7 |
| GLM-5.2 | 13.3 |
| GLM-5.1 | 12.7 |
| GLM-5 | 12.0 |
| Sonnet 4.6 | 10.7 |
| Gemini 2.5 Pro | 5.3 |