GPQA Diamond leaderboard
GPQA Diamond is the hardest tier of the GPQA benchmark — graduate-level science questions designed to be Google-proof.
40 models ranked, highest score first.
| # | Model | Company | Score |
|---|---|---|---|
| 1 | MiniMax-M3 | MiniMax | 92.9% |
| 2 | Claude Fable 5 | Anthropic | 92.6% |
| 3 | Qwen3.7 Max | Alibaba | 92.3% |
| 4 | Claude Opus 4.8 | Anthropic | 92.0% |
| 5 | GPT-5.3 Codex | OpenAI | 91.5% |
| 6 | GPT-5.2 | OpenAI | 90.3% |
| 7 | Qwen3.7 Plus | Alibaba | 90.0% |
| 8 | GPT-5.1 | OpenAI | 87.3% |
| 9 | Claude Opus 4.5 | Anthropic | 86.6% |
| 10 | Ring-2.6-1T | InclusionAI | 85.7% |
| 11 | Gemini 3.1 Pro | 84.2% | |
| 12 | DeepSeek V3.2 | DeepSeek | 84.0% |
| 13 | JT-35B-Flash | China Mobile | 82.9% |
| 14 | Gemini 3.5 Flash | 82.8% | |
| 15 | Step 3.7 Flash | StepFun | 80.9% |
| 16 | Claude Sonnet 4.6 | Anthropic | 79.7% |
| 17 | Claude Opus 4.1 | Anthropic | 79.1% |
| 18 | Gemini 2.5 Flash | 79.0% | |
| 19 | Gemini 3 Pro | 78.5% | |
| 20 | North Mini Code | Cohere | 75.7% |
| 21 | Gemma 4 12B | 75.3% | |
| 22 | Kimi K2 | Moonshot AI | 74.1% |
| 23 | Claude Sonnet 4 | Anthropic | 74.0% |
| 24 | HyperNova 60B 2605 | Multiverse Computing | 73.3% |
| 25 | Claude Sonnet 3.7 | Anthropic | 68.3% |
| 26 | Mistral Large 3 | Mistral | 68.0% |
| 27 | GPT-5 | OpenAI | 67.3% |
| 28 | Gemini 2.0 Flash | 63.6% | |
| 29 | Claude 3.5 Sonnet | Anthropic | 56.0% |
| 30 | DeepSeek-V3 | DeepSeek | 55.7% |
| 31 | LFM2.5-8B-A1B | Liquid AI | 51.3% |
| 32 | Grok-2 | xAI | 51.0% |
| 33 | Claude 3 Opus | Anthropic | 48.9% |
| 34 | Claude 3 Sonnet | Anthropic | 40.0% |
| 35 | Claude 3 Haiku | Anthropic | 37.4% |
| 36 | Gemini 1.5 Pro | 37.1% | |
| 37 | Mistral Large | Mistral | 35.1% |
| 38 | Claude 2.1 | Anthropic | 31.9% |
| 39 | MiniCPM-V 4.6 1.3B | OpenBMB | 30.5% |
| 40 | MiniCPM5-1B | OpenBMB | 27.8% |