Science · %
GPQA Diamond
198 graduate-level multiple-choice questions in biology, physics and chemistry written to be 'Google-proof'.
Active Contamination risk: Medium 282 results · 217 models Updated Oct 22, 2026 Newest result: 2026-10-19 Official leaderboard ↗
What it measures
Expert-level scientific reasoning. Scores here are Epoch AI's own runs (accuracy, best across scorers).
Limitations
Only 198 questions, so ±3 points is noise; random guessing gives 25%; top models are close to saturation.
Contamination risk: Medium
Questions are public since 2023.
Leaderboard
Top 100 of 217
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | Gemini 4 Argon | 96.4% | 1 | Epoch AI run, accuracy; effort: high | Independent run | Oct 13, 2026 | |
| 2 | GPT-6 Astra | OpenAI | 95.8% | 2 | Epoch AI run, accuracy; effort: max | Independent run | Aug 30, 2026 |
| 3 | Claude Sonnet 5.5 | Anthropic | 95.6% | 3 | Epoch AI run, accuracy; effort: max | Independent run | Sep 29, 2026 |
| 4 | GPT-6.1 Sol | OpenAI | 95.4% | 4 | Epoch AI run, accuracy; effort: max | Independent run | Sep 29, 2026 |
| 5 | Gemini 3.8 Flash | 95.4% | 4 | Epoch AI run, accuracy; effort: high | Independent run | Sep 2, 2026 | |
| 6 | Gemini 3.7 Flash | 94.8% | 6 | Epoch AI run, accuracy; effort: high | Independent run | Aug 14, 2026 | |
| 7 | GPT-5.4 Pro | OpenAI | 94.6% | 7 | Epoch AI run, accuracy; effort: xhigh | Independent run | Mar 20, 2026 |
| 8 | Gemini 3.1 Pro | 94.4% | 8 | Epoch AI run, accuracy | Independent run | Aug 6, 2026 | |
| 9 | GPT-6 Sol | OpenAI | 94.3% | 9 | Epoch AI run, accuracy; effort: max | Independent run | Sep 22, 2026 |
| 10 | Grok 4.6 | xAI | 94.0% | 10 | Epoch AI run, accuracy; effort: high | Independent run | Aug 12, 2026 |
| 11 | GPT-5.5 | OpenAI | 94.0% | 10 | Epoch AI run, accuracy; effort: xhigh | Independent run | Apr 24, 2026 |
| 12 | GPT-5.5 Pro | OpenAI | 93.9% | 12 | Epoch AI run, accuracy; effort: xhigh | Independent run | Apr 24, 2026 |
| 13 | Claude Opus 5 | Anthropic | 93.9% | 13 | Epoch AI run, accuracy; effort: max | Independent run | Jul 24, 2026 |
| 14 | GPT-5.6 Sol | OpenAI | 93.5% | 14 | Epoch AI run, accuracy; effort: max | Independent run | Jul 9, 2026 |
| 15 | Grok 4.5 | xAI | 93.4% | 15 | Epoch AI run, accuracy; effort: high | Independent run | Jul 8, 2026 |
| 16 | GPT-5.6 Terra | OpenAI | 93.3% | 16 | Epoch AI run, accuracy; effort: max | Independent run | Jul 9, 2026 |
| 17 | GPT-5.4 | OpenAI | 93.3% | 17 | Epoch AI run, accuracy; effort: xhigh | Independent run | Mar 6, 2026 |
| 18 | DeepSeek V4.1 Flash 1017OW | DeepSeek | 93.2% | 18 | Epoch AI run, accuracy; effort: max | Independent run | Oct 19, 2026 |
| 19 | Kimi K3OW | Moonshot AI | 93.1% | 20 | Epoch AI run, accuracy; effort: max | Independent run | Jul 16, 2026 |
| 20 | Grok 4.7 | xAI | 92.7% | 22 | Epoch AI run, accuracy; effort: xhigh | Independent run | Sep 22, 2026 |
| 21 | Qwen 3.8 Max | Alibaba (Qwen) | 92.7% | 23 | Epoch AI run, accuracy; effort: xhigh | Independent run | Aug 4, 2026 |
| 22 | Gemini 3 Pro | 92.6% | 24 | Epoch AI run, accuracy | Independent run | Nov 19, 2025 | |
| 23 | Qwen3.8 Max (0902) | Alibaba (Qwen) | 92.3% | 25 | Epoch AI run, accuracy; effort: xhigh | Independent run | Sep 2, 2026 |
| 24 | DeepSeek V4 Pro 0813OW | DeepSeek | 91.7% | 27 | Epoch AI run, accuracy; effort: max | Independent run | Aug 18, 2026 |
| 25 | GPT-5.6 Luna | OpenAI | 91.6% | 28 | Epoch AI run, accuracy; effort: max | Independent run | Jul 9, 2026 |
| 26 | GPT-5.2 | OpenAI | 91.4% | 29 | Epoch AI run, accuracy; effort: xhigh | Independent run | Dec 13, 2025 |
| 27 | DeepSeek V4 Flash 0731OW | DeepSeek | 91.0% | 30 | Epoch AI run, accuracy; effort: max | Independent run | Aug 2, 2026 |
| 28 | GLM-5.3OW | Zhipu AI (Z.ai) | 90.9% | 31 | Epoch AI run, accuracy; effort: max | Independent run | Aug 24, 2026 |
| 29 | Qwen3.7-Max | Alibaba (Qwen) | 90.9% | 31 | Epoch AI run, accuracy | Independent run | Aug 7, 2026 |
| 30 | DeepSeek-V4-ProOW | DeepSeek | 90.9% | 31 | Epoch AI run, accuracy; effort: high | Independent run | Aug 6, 2026 |
| 31 | Kimi K2.6OW | Moonshot AI | 90.8% | 34 | Epoch AI run, accuracy | Independent run | May 1, 2026 |
| 32 | Claude Opus 5.5 | Anthropic | 90.6% | 36 | Epoch AI run, accuracy; effort: max | Independent run | Sep 22, 2026 |
| 33 | Claude Sonnet 5 | Anthropic | 90.5% | 37 | Epoch AI run, accuracy; effort: xhigh | Independent run | Jul 1, 2026 |
| 34 | Claude Opus 4.6 | Anthropic | 90.5% | 37 | Epoch AI run, accuracy; effort: 32k thinking | Independent run | Feb 6, 2026 |
| 35 | GPT-6 Luna | OpenAI | 90.5% | 39 | Epoch AI run, accuracy; effort: max | Independent run | Sep 22, 2026 |
| 36 | GLM-5.3-FlashOW | Zhipu AI (Z.ai) | 90.2% | 40 | Epoch AI run, accuracy; effort: max | Independent run | Aug 26, 2026 |
| 37 | Claude Opus 4.7 | Anthropic | 90.2% | 40 | Epoch AI run, accuracy; effort: xhigh | Independent run | Apr 17, 2026 |
| 38 | GLM-5.1OW | Zhipu AI (Z.ai) | 89.9% | 42 | Epoch AI run, accuracy | Independent run | Aug 10, 2026 |
| 39 | DeepSeek V4.1 FlashOW | DeepSeek | 89.8% | 45 | Epoch AI run, accuracy; effort: max | Independent run | Oct 9, 2026 |
| 40 | Muse Spark | Meta | 89.8% | 46 | Epoch AI run, accuracy | Independent run | Apr 8, 2026 |
| 41 | Claude Haiku 5.5 | Anthropic | 89.7% | 47 | Epoch AI run, accuracy; effort: max | Independent run | Oct 9, 2026 |
| 42 | Gemini 3 Flash | 89.4% | 49 | Epoch AI run, accuracy | Independent run | Aug 6, 2026 | |
| 43 | Grok 4.20 | xAI | 89.3% | 50 | Epoch AI run, accuracy | Independent run | Jul 13, 2026 |
| 44 | Gemini 3.5 Flash | 88.9% | 51 | Epoch AI run, accuracy | Independent run | Aug 6, 2026 | |
| 45 | Grok 4.3 Beta | xAI | 88.8% | 53 | Epoch AI run, accuracy | Independent run | Jun 17, 2026 |
| 46 | Muse Spark OpenOW | Meta | 88.6% | 55 | Epoch AI run, accuracy; effort: xhigh | Independent run | Oct 15, 2026 |
| 47 | Inkling-SmallOW | Thinking Machines | 88.5% | 56 | Epoch AI run, accuracy; effort: xhigh | Independent run | Aug 14, 2026 |
| 48 | Qwen 3.6 Plus | Alibaba (Qwen) | 88.4% | 57 | Epoch AI run, accuracy; effort: 2026-04-02 | Independent run | Aug 7, 2026 |
| 49 | InklingOW | Thinking Machines | 88.3% | 59 | Epoch AI run, accuracy; effort: xhigh | Independent run | Aug 5, 2026 |
| 50 | Mistral Large 4 | Mistral AI | 88.1% | 61 | Epoch AI run, accuracy; effort: high | Independent run | Oct 12, 2026 |
| 51 | Kimi K2.7 CodeOW | Moonshot AI | 87.9% | 62 | Epoch AI run, accuracy | Independent run | Aug 7, 2026 |
| 52 | Qwen3.7-Plus | Alibaba (Qwen) | 87.9% | 62 | Epoch AI run, accuracy | Independent run | Aug 7, 2026 |
| 53 | GLM-5OW | Zhipu AI (Z.ai) | 87.8% | 66 | Epoch AI run, accuracy | Independent run | Feb 12, 2026 |
| 54 | GPT-5.1 | OpenAI | 87.6% | 67 | Epoch AI run, accuracy; effort: high | Independent run | Nov 13, 2025 |
| 55 | Kimi K2.5OW | Moonshot AI | 87.6% | 68 | Epoch AI run, accuracy; effort: Fireworks | Independent run | Feb 2, 2026 |
| 56 | Qwen 3.6 Max (Preview) | Alibaba (Qwen) | 87.4% | 69 | Epoch AI run, accuracy; effort: Preview | Independent run | Aug 7, 2026 |
| 57 | Claude Sonnet 4.6 | Anthropic | 87.4% | 69 | Epoch AI run, accuracy; effort: 32k thinking | Independent run | Feb 20, 2026 |
| 58 | Grok 4 | xAI | 87.0% | 72 | Epoch AI run, accuracy | Independent run | Jul 9, 2025 |
| 59 | GPT-5.4 Mini | OpenAI | 86.9% | 73 | Epoch AI run, accuracy; effort: xhigh | Independent run | Aug 7, 2026 |
| 60 | GPT-5 | OpenAI | 86.2% | 75 | Epoch AI run, accuracy; effort: high | Independent run | Oct 29, 2025 |
| 61 | Claude Opus 4.5 | Anthropic | 86.0% | 76 | Epoch AI run, accuracy; effort: 32k thinking | Independent run | Nov 24, 2025 |
| 62 | Qwen3.6 27BOW | Alibaba (Qwen) | 85.9% | 77 | Epoch AI run, accuracy | Independent run | Aug 7, 2026 |
| 63 | Qwen3.5 397B-A17BOW | Alibaba (Qwen) | 85.9% | 77 | Epoch AI run, accuracy | Independent run | Aug 7, 2026 |
| 64 | Gemini 3.6 Flash | 85.9% | 77 | Epoch AI run, accuracy | Independent run | Aug 7, 2026 | |
| 65 | Claude Fable 5 | Anthropic | 85.9% | 77 | Epoch AI run, accuracy; effort: max | Independent run | Aug 6, 2026 |
| 66 | Nemotron 3 UltraOW | NVIDIA | 85.3% | 82 | Epoch AI run, accuracy | Independent run | Aug 10, 2026 |
| 67 | Claude Opus 4.8 | Anthropic | 85.3% | 82 | Epoch AI run, accuracy | Independent run | Aug 6, 2026 |
| 68 | Gemini 2.5 Pro (Jun 2025) | 85.3% | 85 | Epoch AI run, accuracy | Independent run | Nov 16, 2025 | |
| 69 | Qwen 3.5 Plus (hosted 397B-A17B) | Alibaba (Qwen) | 84.8% | 87 | Epoch AI run, accuracy | Independent run | Aug 7, 2026 |
| 70 | Kimi K2 ThinkingOW | Moonshot AI | 84.2% | 91 | Epoch AI run, accuracy | Independent run | Nov 11, 2025 |
| 71 | Qwen 3.6 35B-A3BOW | Alibaba (Qwen) | 83.8% | 92 | Epoch AI run, accuracy | Independent run | Aug 7, 2026 |
| 72 | Gemini 2.5 Pro (Mar 2025) | 83.8% | 92 | Epoch AI run, accuracy; effort: Mar 2025 | Independent run | Mar 31, 2025 | |
| 73 | DeepSeek-V3.2OW | DeepSeek | 83.4% | 95 | Epoch AI run, accuracy; effort: Thinking | Independent run | Dec 16, 2025 |
| 74 | Qwen 3.6 Flash | Alibaba (Qwen) | 83.3% | 96 | Epoch AI run, accuracy; effort: 2026-04-16 | Independent run | Aug 7, 2026 |
| 75 | GLM-4.7OW | Zhipu AI (Z.ai) | 83.3% | 96 | Epoch AI run, accuracy | Independent run | Jan 29, 2026 |
| 76 | GPT-5.5 Instant | OpenAI | 82.5% | 103 | Epoch AI run, accuracy | Independent run | Aug 2, 2026 |
| 77 | Qwen3.7 Flash | Alibaba (Qwen) | 82.3% | 104 | Epoch AI run, accuracy | Independent run | Aug 7, 2026 |
| 78 | Qwen 3.5 Flash (hosted 35B-A3B) | Alibaba (Qwen) | 82.3% | 104 | Epoch AI run, accuracy | Independent run | Aug 7, 2026 |
| 79 | Claude Sonnet 4.5 | Anthropic | 82.3% | 104 | Epoch AI run, accuracy; effort: 59k thinking | Independent run | Oct 28, 2025 |
| 80 | o3 | OpenAI | 81.8% | 108 | Epoch AI run, accuracy; effort: high | Independent run | Apr 16, 2025 |
| 81 | MiniMax-M3OW | MiniMax | 81.3% | 110 | Epoch AI run, accuracy | Independent run | Aug 10, 2026 |
| 82 | Qwen3.5-35B-A3BOW | Alibaba (Qwen) | 81.2% | 111 | Epoch AI run, accuracy | Independent run | Aug 28, 2026 |
| 83 | Qwen3-235B-A22B-Thinking (Jul 2025)OW | Alibaba (Qwen) | 80.0% | 115 | Epoch AI run, accuracy | Independent run | Dec 11, 2025 |
| 84 | Claude 3.7 Sonnet | Anthropic | 79.7% | 117 | Epoch AI run, accuracy; effort: 64k thinking | Independent run | May 26, 2025 |
| 85 | o4-mini | OpenAI | 79.6% | 118 | Epoch AI run, accuracy; effort: high | Independent run | Apr 16, 2025 |
| 86 | Claude Sonnet 4 | Anthropic | 79.2% | 119 | Epoch AI run, accuracy | Independent run | May 26, 2025 |
| 87 | Qwen3.5-9BOW | Alibaba (Qwen) | 78.9% | 120 | Epoch AI run, accuracy | Independent run | Aug 27, 2026 |
| 88 | GPT-5.4 Nano | OpenAI | 78.5% | 124 | Epoch AI run, accuracy; effort: high | Independent run | Apr 14, 2026 |
| 89 | o3-mini | OpenAI | 77.0% | 128 | Epoch AI run, accuracy; effort: high | Independent run | Feb 13, 2025 |
| 90 | Claude Opus 4.1 | Anthropic | 76.8% | 129 | Epoch AI run, accuracy | Independent run | Aug 5, 2025 |
| 91 | o1 | OpenAI | 76.8% | 129 | Epoch AI run, accuracy; effort: high | Independent run | Feb 13, 2025 |
| 92 | DeepSeek-R1 (May 2025)OW | DeepSeek | 76.3% | 133 | Epoch AI run, accuracy; effort: May 2025 | Independent run | May 29, 2025 |
| 93 | Claude Opus 4 | Anthropic | 76.3% | 134 | Epoch AI run, accuracy | Independent run | May 22, 2025 |
| 94 | Gemma 4 31B ITOW | 75.8% | 135 | Epoch AI run, accuracy | Independent run | Aug 6, 2026 | |
| 95 | Gemini 3.5 Flash-Lite | 75.8% | 135 | Epoch AI run, accuracy | Independent run | Aug 6, 2026 | |
| 96 | gpt-oss-120bOW | OpenAI | 75.8% | 135 | Epoch AI run, accuracy | Independent run | Dec 11, 2025 |
| 97 | Grok 3 | xAI | 75.8% | 135 | Epoch AI run, accuracy | Independent run | May 26, 2025 |
| 98 | Grok-3 mini | xAI | 75.5% | 140 | Epoch AI run, accuracy | Independent run | May 26, 2025 |
| 99 | GPT-5 mini | OpenAI | 75.0% | 142 | Epoch AI run, accuracy; effort: high | Independent run | Oct 30, 2025 |
| 100 | Gemini 3.1 Flash-Lite | 73.7% | 146 | Epoch AI run, accuracy | Independent run | Aug 6, 2026 |
Best score over time
Each step is a new best result in our data, by evaluation date.
- 2025-01-2775.8%o1
- 2025-02-1377.0%o3-mini
- 2025-03-3183.8%Gemini 2.5 Pro (Mar 2025)
- 2025-06-0584.8%Gemini 2.5 Pro (Jun 2025)
- 2025-07-0987.0%Grok 4
- 2025-11-1387.6%GPT-5.1
- 2025-11-1992.6%Gemini 3 Pro
- 2026-03-0693.3%GPT-5.4
- 2026-03-2094.6%GPT-5.4 Pro
- 2026-08-1494.8%Gemini 3.7 Flash
- 2026-08-3095.8%GPT-6 Astra
- 2026-10-1396.4%Gemini 4 Argon
Data sources & licences
| Source | Licence / terms | How we reuse it | Last fetched |
|---|---|---|---|
| Epoch AI Benchmarking Hub Epoch AI, 'Capabilities & benchmarking', https://epoch.ai/benchmarks (CC BY 4.0) | CC BY 4.0 | Open licence – full data with attribution | Oct 22, 2026 |
FAQ
What does GPQA Diamond measure?
198 graduate-level multiple-choice questions in biology, physics and chemistry written to be 'Google-proof'. Expert-level scientific reasoning. Scores here are Epoch AI's own runs (accuracy, best across scorers).
How reliable is GPQA Diamond?
Only 198 questions, so ±3 points is noise; random guessing gives 25%; top models are close to saturation. Contamination risk: Medium – Questions are public since 2023.