GPQA Diamond
Epoch AI / Knowledge · Graduate-level physics, chemistry and biology knowledge
What it measures, and how
Only Epoch's own re-run mean_score is used, retaining the reasoning tier and standard error. The 198 science knowledge questions require reasoning but do not represent all world knowledge.
How this evidence is used
Officially scored; the same question family shares a fixed budget.
Results
Under fixed rules, each public model uses one representative configuration. Anonymous test variants are not shown, original leaderboard ranks are kept, and each item lists at most 30.
| Rank there | Model there | Raw score | Representative configuration |
|---|---|---|---|
| 1 | gpt-6-astra_maxOpenAI | 95.8% | Max reasoning |
| 2 | claude-sonnet-5-5_maxanthropic | 95.6% | Max reasoning |
| 3 | gpt-6.1-sol_max— | 95.4% | Max reasoning |
| 3 | gemini-3.8-flash_highGoogle | 95.4% | High reasoning |
| 5 | gemini-3.7-flash_highGoogle | 94.8% | High reasoning |
| 6 | gpt-5.4-pro-2026-03-05_xhighOpenAI | 94.6% | xHigh reasoning |
| 7 | gemini-3.1-pro-preview_highGoogle | 94.4% | High reasoning |
| 8 | gpt-6-sol_maxOpenAI | 94.3% | Max reasoning |
| 9 | gemini-3.6-flash_highGoogle | 94.1% | High reasoning |
| 14 | claude-opus-5_maxAnthropic | 93.9% | Max reasoning |
| 15 | gpt-5.6-sol_maxOpenAI | 93.5% | Max reasoning |
| 16 | grok-4.5_highxAI | 93.4% | High reasoning |
| 17 | gpt-5.6-terra_maxOpenAI | 93.3% | Max reasoning |
| 18 | gpt-5.4-2026-03-05_xhighOpenAI | 93.3% | xHigh reasoning |
| 19 | grok-4.6_xhighxAI | 93.2% | xHigh reasoning |
| 20 | kimi-k3_maxMoonshot AI | 93.1% | Max reasoning |
| 22 | gemini-3.5-flash_highGoogle | 92.8% | High reasoning |
| 23 | qwen3.8-max_xhighAlibaba | 92.7% | xHigh reasoning |
| 24 | gemini-3-pro-previewGoogle | 92.6% | Source default configuration |
| 25 | qwen3.8-max-0902_xhighAlibaba | 92.3% | xHigh reasoning |
| 27 | glm-5.2_maxZ.ai | 91.9% | Max reasoning |
| 28 | deepseek-v4-pro-0813_maxDeepSeek | 91.7% | Max reasoning |
| 29 | gpt-5.6-luna_maxOpenAI | 91.6% | Max reasoning |
| 30 | gpt-5.2-2025-12-11_xhighOpenAI | 91.4% | xHigh reasoning |
| 31 | claude-opus-4-8_maxAnthropic | 91.0% | Max reasoning |
| 31 | deepseek-v4-flash-0731_maxDeepSeek | 91.0% | Max reasoning |
| 33 | glm-5.3_maxZ.ai | 90.9% | Max reasoning |
| 33 | MiniMax-M3MiniMax | 90.9% | Source default configuration |
| 33 | qwen3.7-maxAlibaba | 90.9% | Source default configuration |
| 37 | kimi-k2.6Moonshot AI | 90.8% | Source default configuration |
Limits and data attribution
Officially scored; the same question family shares a fixed budget.
Data licence: CC BY 4.0 · Epoch AI's own evaluation data
Scores published by Epoch AI. Raw scores and the News consensus score use different scales and cannot be added directly.