Reasoning · weight 15%推理 · 权重 15%
GPQA Diamond
What it tests: Graduate-level multiple-choice science questions (biology, physics, chemistry) written so that experts with web access still find them hard. 'Diamond' is the highest-quality 198-question subset.
考什么:研究生级别的科学选择题(生物/物理/化学),专门设计成'专家带着搜索引擎也觉得难'。Diamond 是其中质量最高的 198 题子集。
Why it counts: One of the last knowledge benchmarks where frontier models still make mistakes — it separates genuine scientific reasoning from memorized answers.
为什么算数:少数还能让顶尖模型犯错的 benchmark 之一——区分真科学推理和背答案。头部模型约 80-96%,尚未触顶。
| Maintainer维护方 | NYU / Anthropic 等联合 |
|---|---|
| License许可 | 学术免费,商用需授权 |
| Reference point (prior-gen SOTA)参考点(上一代 SOTA) | 80 |
◈
Standings实时排名
60 models with sourced scores 个模型有溯源成绩| # | Model模型 | Score成绩 | Index指数 | Dated日期 | Measured by测评方 | |
|---|---|---|---|---|---|---|
| 1 | Gemini 3.1 Pro PreviewGoogle | 95.5 | 100 | — | 3rd-party第三方 | source ↗×9 |
| 2 | Claude Opus 4.8Anthropic | 94.3 | 99 | 2026-07-01 | 3rd-party第三方 | source ↗×4 |
| 3 | Claude Fable 5Anthropic | 94.1⚡max | 99 | — | 3rd-party第三方 | source ↗×8⚠±15.7 |
| 4 | GPT-5.6 SolOpenAI | 94.1⚡max | 99 | — | 3rd-party第三方 | source ↗×7 |
| 5 | GPT-5.6 Sol ProOpenAI | 93.8 | 98 | — | 3rd-party第三方 | source ↗ |
| 6 | Claude Opus 5Anthropic | 93.7 | 98 | — | 3rd-party第三方 | source ↗×9⚠±5.2 |
| 7 | Kimi K3Moonshot AI | 93.5 | 98 | 2026-07-18 | 3rd-party第三方 | source ↗×7 |
| 8 | GPT-5.5OpenAI | 93.5⚡high | 98 | — | 3rd-party第三方 | source ↗×7 |
| 9 | Gemini 3.6 FlashGoogle | 93.43 | 98 | 2026-08 | 3rd-party第三方 | source ↗×5 |
| 10 | Grok 4.5xAI | 92.9 | 97 | 2026-07-01 | 3rd-party第三方 | source ↗×5⚠±6.5 |
| 11 | Qwen3.8 Max (0902)Qwen | 92.6 | 97 | — | 3rd-party第三方 | source ↗×4 |
| 12 | Qwen3.7 MaxQwen | 92.4 | 97 | — | 3rd-party第三方 | source ↗×6 |
| 13 | Gemini 3.5 FlashGoogle | 92.2⚡high | 97 | — | 3rd-party第三方 | source ↗×5 |
| 14 | GLM 5.2Z.ai | 91.2 | 96 | 2026-07-18 | 3rd-party第三方 | source ↗×8⚠±3.8 |
| 15 | MiniMax M3MiniMax | 91.2 | 96 | — | 3rd-party第三方 | source ↗×4 |
| 16 | GPT-5.6 TerraOpenAI | 91.1⚡high | 95 | — | 3rd-party第三方 | source ↗×5⚠±3.8 |
| 17 | GPT-5.6 LunaOpenAI | 91.1⚡max | 95 | — | 3rd-party第三方 | source ↗×3 |
| 18 | Grok 4.20xAI | 91 | 95 | — | 3rd-party第三方 | source ↗×4⚠±3.5 |
| 19 | DeepSeek V4 Flash 0731DeepSeek | 90.8 | 95 | — | 3rd-party第三方 | source ↗×2 |
| 20 | Hy3Tencent | 90.4 | 95 | 2026-08 | 3rd-party第三方 | source ↗×2 |
| 21 | Muse Spark 1.2meta | 90.4 | 95 | — | 3rd-party第三方 | source ↗ |
| 22 | Hunyuan A13B InstructTencent | 90.4 | 95 | — | 3rd-party第三方 | source ↗×2⚠±7.7 |
| 23 | Qwen3.7 PlusQwen | 90.3 | 95 | 2026-06 | 3rd-party第三方 | source ↗×4 |
| 24 | Qwen3.7 FlashQwen | 90.15 | 94 | — | 3rd-party第三方 | source ↗×2 |
| 25 | DeepSeek V4 Pro 0423DeepSeek | 90.1 | 94 | 2026-07-18 | 3rd-party第三方 | source ↗×8⚠±4.7 |
| 26 | Grok 4.3xAI | 90.1⚡high | 94 | — | 3rd-party第三方 | source ↗×4 |
| 27 | GPT-5.6 Luna ProOpenAI | 89.5 | 94 | — | 3rd-party第三方 | source ↗ |
| 28 | Muse Spark 1.1meta | 89.5 | 94 | 2026-08 | 3rd-party第三方 | source ↗×2 |
| 29 | DeepSeek V4 Flash 0423DeepSeek | 87.4 | 92 | — | 3rd-party第三方 | source ↗×4⚠±3.6 |
| 30 | Nemotron 3 UltraNVIDIA | 87 | 91 | — | 3rd-party第三方 | source ↗×2 |
| 31 | MiMo-V2.5-ProXiaomi | 87 | 91 | — | 3rd-party第三方 | source ↗×2 |
| 32 | GLM 5.1Z.ai | 87⚡high | 91 | — | 3rd-party第三方 | source ↗×5⚠±3.7 |
| 33 | Gemini 3.1 Flash LiteGoogle | 86.9 | 91 | — | 3rd-party第三方 | source ↗×5 |
| 34 | MiniMax M2.7MiniMax | 86.6 | 91 | 2026-03-01 | 3rd-party第三方 | source ↗×6⚠±5.4 |
| 35 | Seed-2.0-LiteByteDance Seed | 85.1 | 89 | — | 3rd-party第三方 | source ↗ |
| 36 | Kimi K2 ThinkingMoonshot AI | 84.5⚡high | 88 | — | 3rd-party第三方 | source ↗×6 |
| 37 | Gemini 3.5 Flash LiteGoogle | 84 | 88 | — | 3rd-party第三方 | source ↗×3⚠±3.1 |
| 38 | DeepSeek V3.2DeepSeek | 82.9 | 87 | — | 3rd-party第三方 | source ↗×4⚠±5.4 |
| 39 | gpt-oss-120bOpenAI | 80.1⚡high | 84 | — | 3rd-party第三方 | source ↗×2 |
| 40 | Seed-2.0-MiniByteDance Seed | 79 | 83 | — | 3rd-party第三方 | source ↗×2 |
| 41 | LongCat 2.0Meituan | 78 | 82 | — | 3rd-party第三方 | source ↗×2⚠±10.9 |
| 42 | Kimi K2.7 CodeMoonshot AI | 77.8 | 81 | — | 3rd-party第三方 | source ↗×5⚠±12.2 |
| 43 | Mercury 2Inception | 77 | 81 | — | 3rd-party第三方 | source ↗×2 |
| 44 | Step 3.7 FlashStepFun | 76.6 | 80 | — | 3rd-party第三方 | source ↗×3⚠±4.3 |
| 45 | Mistral Medium 3.5Mistral AI | 74.8 | 78 | — | 3rd-party第三方 | source ↗×4 |
| 46 | Claude Sonnet 5Anthropic | 74.6 | 78 | — | 3rd-party第三方 | source ↗×6⚠±21.6 |
| 47 | ERNIE 4.5 VL 424B A47B Baidu | 74 | 77 | — | 3rd-party第三方 | source ↗ |
| 48 | Sonar ProPerplexity | 72.4 | 76 | — | 3rd-party第三方 | source ↗×2⚠±10.1 |
| 49 | Solar Pro 3Upstage | 72.4 | 76 | — | 3rd-party第三方 | source ↗ |
| 50 | Llama 4 MaverickMeta | 69.8 | 73 | — | 3rd-party第三方 | source ↗×7 |
| 51 | Claude Haiku 4.5Anthropic | 64.6 | 68 | — | 3rd-party第三方 | source ↗×4⚠±8.4 |
| 52 | Llama 4 ScoutMeta | 57.2 | 60 | — | 3rd-party第三方 | source ↗×5 |
| 53 | Nova Premier 1.0Amazon | 56.9 | 60 | — | 3rd-party第三方 | source ↗×3 |
| 54 | Phi 4Microsoft | 56.1 | 59 | — | 3rd-party第三方 | source ↗×3⚠±13.8 |
| 55 | Command ACohere | 50.8 | 53 | — | 3rd-party第三方 | source ↗×3⚠±25.3 |
| 56 | Nova Pro 1.0Amazon | 49.9 | 52 | — | 3rd-party第三方 | source ↗×2 |
| 57 | Palmyra X5Writer | 47.2 | 49 | — | vendor-reported厂商自报 | source ↗ |
| 58 | Mistral Small 4Mistral AI | 44.2 | 46 | 2026-04-01 | 3rd-party第三方 | source ↗×2⚠±6.1 |
| 59 | Mistral LargeMistral AI | 43.9 | 46 | 2026 | 3rd-party第三方 | source ↗×6⚠±24.1 |
| 60 | Command R+ (08-2024)Cohere | 31.1 | 33 | — | 3rd-party第三方 | source ↗×2 |