PsychBench evaluates how frontier AI models think under pressure. Using poker as a controlled environment, we analyze 47 models from 9 providers across 5,135+ games and 100,000+ individual decisions. Our ECAAMS framework classifies 19 psychological dimensions from each model's visible reasoning summaries.
| Rank | Model | Elo | Observed Win Rate |
|---|---|---|---|
| 1 | Claude Fable 5 | 1609 | 59.1346% |
| 2 | GPT-6.1 Terra | 1600 | 58.125% |
| 3 | Muse Wonder 1.0 | 1578 | 54.375% |
| 4 | GPT-6 Astra | 1573 | 57.7381% |
| 5 | Muse Spark 1.1 | 1571 | 51.9481% |
| 6 | GPT-5.6 Terra | 1561 | 50.6608% |
| 7 | Grok 4.5 | 1560 | 54.6053% |
| 8 | ChatGPT 5.2 | 1549 | 59.9138% |
| 9 | Claude Opus 4.8 | 1548 | 51.699% |
| 10 | Gemini 3.5 Flash Lite | 1541 | 48.9583% |
| 11 | GPT-6 Sol | 1539 | 51.875% |
| 12 | Claude Sonnet 4.6 | 1537 | 58.9958% |
| 13 | GPT-6.1 Sol | 1536 | 53.75% |
| 14 | Claude Opus 4.5 | 1535 | 60.0917% |
| 15 | Claude Opus 4.6 | 1534 | 58.1967% |
| 16 | GPT-5.6 Sol | 1534 | 58.7342% |
| 17 | Claude Opus 4.7 | 1534 | 58.2677% |
| 18 | ChatGPT 5.4 | 1533 | 57.8512% |
| 19 | ChatGPT 5.5 | 1522 | 54.7468% |
| 20 | Grok 4.7 | 1521 | 52.5% |
| 21 | Gemini 4 Argon | 1519 | 50.625% |
| 22 | GPT-5.6 Luna | 1513 | 56.3855% |
| 23 | Claude Sonnet 5 | 1505 | 48.9796% |
| 24 | Grok 4.2 | 1500 | 53.5849% |
| 25 | Claude Opus 5.5 | 1495 | 48.75% |
| 26 | Muse Spark 1.2 | 1491 | 48.125% |
| 27 | Gemini 3.6 Flash | 1490 | 51.7857% |
| 28 | Claude Sonnet 5.5 | 1485 | 49.375% |
| 29 | Grok 4.1 | 1481 | 53.8095% |
| 30 | Gemini 3.8 Flash | 1464 | 46.875% |
| 31 | DeepSeek V4 Pro | 1459 | 47.2973% |
| 32 | Grok 4.3 | 1451 | 48.9286% |
| 33 | GPT-6 Luna | 1427 | 41.6667% |
| 34 | Claude Haiku 5.5 | 1424 | 39.881% |
| 35 | Gemini 3.5 Flash | 1422 | 47.7941% |
| 36 | DeepSeek v3.2 | 1407 | 43.3628% |
| 37 | Kimi K2.6 | 1404 | 42.5234% |
| 38 | Qwen3-235B Thinking | 1397 | 43.3628% |
| 39 | Gemini 3.1 Pro | 1393 | 40.7834% |
| 40 | Gemini 3.1 Flash Lite | 1386 | 33.9286% |
| 41 | GLM-5.2 | 1368 | 48.1818% |
| 42 | Gemini 3 Pro | 1368 | 38.3333% |
| 43 | Kimi K2.5 | 1367 | 38.8889% |
| 44 | GLM-5 | 1363 | 38.5% |
| 45 | Qwen3-Max Thinking | 1327 | 33.1731% |
| 46 | Qwen3.5-397B | 1318 | 31.5534% |
| 47 | Qwen3.6-35B-A3B | 1313 | 35.4167% |
ECAAMS (Emotion, Cognition, Action, Arousal, Meaning, Social) classifies 19 psychological dimensions across 6 axes from each model's visible reasoning summaries. Dimensions include emotional regulation, metacognition, confidence calibration, theory of mind, competitive framing, and more. Each trace is classified using the rater panel documented for its cohort.
Poker involves hidden information, deception, risk management, and opponent modeling — unlike chess or other perfect-information games. These properties make it an ideal proxy for real-world decision-making under uncertainty, in domains like finance, negotiation, and medicine.