Mystery Game Puzzles
Games · 2026-07-17
Epoch AI benchmark of 100 positions from a puzzle-oriented variant of a well-known game whose identity Epoch keeps secret to deter benchmark-specific preparation. Like Chess Puzzles, each position has a single best next move, which the model submits through a submission tool; an illegal or unextractable move gets no credit. Epoch does not publish the prompt, example positions or transcripts. Atlas pools task versions 1.0.4, 1.1.0 and 1.2.x, which Epoch publishes in one table without a changelog.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 84.0 |
| GPT-6.1 Sol | 80.0 |
| Claude Opus 5.5 | 71.0 |
| Claude Sonnet 5.5 | 65.0 |
| Gemini 4 Argon | 64.0 |
| Claude Opus 5 | 59.0 |
| Claude Fable 5.1 | 58.0 |
| GPT-5.6 Sol | 58.0 |
| GPT-5.5 | 56.0 |
| GPT-6 Sol | 56.0 |
| Claude Fable 5 | 52.0 |
| Muse Wonder | 51.0 |
| DeepSeek V4.1 Flash 1017 | 49.0 |
| Gemini 3.8 Flash | 47.0 |
| DeepSeek V4.1 Flash | 43.0 |
| DeepSeek V4 Pro 0813 | 43.0 |
| Qwen3.8-Max | 38.0 |
| Gemini 3.7 Flash | 37.0 |
| GPT-5.4 | 37.0 |
| Opus 4.8 | 36.0 |
| Sonnet 5 | 35.0 |
| GPT-5.6 Terra | 35.0 |
| Gemini 3.1 Pro Preview | 34.0 |
| Grok 4.6 | 34.0 |
| GLM-5.3 | 33.0 |
| Gemini 3.5 Flash | 32.0 |
| Qwen3.7-Max (source unspecified) | 32.0 |
| Claude Haiku 5.5 | 30.0 |
| Gemini 3.6 Flash | 30.0 |
| Grok 4.7 | 29.0 |
| o3 | 29.0 |
| Opus 4.7 | 28.0 |
| Gemini 3 Flash Preview | 26.0 |
| Kimi K3 | 26.0 |
| Opus 4.6 | 25.0 |
| Muse Spark 1.3 | 25.0 |
| GPT-5.2 | 23.0 |
| GPT-5 | 23.0 |
| Opus 4.5 | 22.0 |
| Qwen3.6 35B A3B | 22.0 |