Knowledge · %
Humanity's Last Exam
About 2,500 expert-written questions across 100+ subjects from the Center for AI Safety and Scale AI, filtered to stump frontier models at creation time.
What it measures
Accuracy (mostly exact-match answers). Data compiled by Epoch AI.
Limitations
Independent audits reported incorrect reference answers in some subjects; settings (tools, text-only) differ between reports.
Contamination risk: Low
Private held-out set exists; public questions carry a canary string.
Leaderboard
46 models
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | GPT-6 Astra | OpenAI | 54.8% | 1 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | Sep 3, 2026 |
| 2 | Claude Fable 5.1 | Anthropic | 46.5% | 2 | Compiled by Epoch AI; effort: xhigh | Third-party compilation via Epoch AI | Sep 1, 2026 |
| 3 | Gemini 3.1 Pro | 46.4% | 3 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Feb 19, 2026 | |
| 4 | Claude Sonnet 5.5 | Anthropic | 46.1% | 4 | Compiled by Epoch AI; effort: xhigh | Third-party compilation via Epoch AI | Sep 28, 2026 |
| 5 | Gemini 3.8 Flash | 44.5% | 5 | Compiled by Epoch AI; effort: unknown | Third-party compilation via Epoch AI | Sep 2, 2026 | |
| 6 | GPT-5.4 Pro | OpenAI | 44.3% | 6 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Mar 5, 2026 |
| 7 | Muse Spark | Meta | 40.6% | 7 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Apr 8, 2026 |
| 8 | Gemini 3 Pro | 37.5% | 8 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Nov 18, 2025 | |
| 9 | GPT-5.4 | OpenAI | 36.2% | 9 | Compiled by Epoch AI; effort: xhigh | Third-party compilation via Epoch AI | Mar 5, 2026 |
| 10 | Claude Opus 4.7 | Anthropic | 36.2% | 10 | Compiled by Epoch AI; effort: unknown | Third-party compilation via Epoch AI | Apr 16, 2026 |
| 11 | Claude Opus 4.6 | Anthropic | 34.4% | 11 | Compiled by Epoch AI; effort: max | Third-party compilation via Epoch AI | Feb 5, 2026 |
| 12 | GPT-5 Pro | OpenAI | 31.6% | 12 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Oct 7, 2025 |
| 13 | Claude Haiku 5.5 | Anthropic | 30.1% | 13 | Compiled by Epoch AI; effort: unknown | Third-party compilation via Epoch AI | Oct 7, 2026 |
| 14 | GPT-5.2 | OpenAI | 27.8% | 14 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | Dec 11, 2025 |
| 15 | GPT-5 | OpenAI | 25.3% | 15 | Compiled by Epoch AI; effort: high | Third-party compilation via Epoch AI | Aug 7, 2025 |
| 16 | Claude Opus 4.5 | Anthropic | 25.2% | 17 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | Nov 24, 2025 |
| 17 | Kimi K2.5OW | Moonshot AI | 24.4% | 18 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Jan 27, 2026 |
| 18 | GPT-5.1 | OpenAI | 23.7% | 19 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | Nov 13, 2025 |
| 19 | Gemini 2.5 Pro (Jun 2025) | 21.6% | 20 | Compiled by Epoch AI; effort: Jun 2025 | Third-party compilation via Epoch AI | Jun 5, 2025 | |
| 20 | o3 | OpenAI | 20.3% | 21 | Compiled by Epoch AI; effort: high | Third-party compilation via Epoch AI | Apr 16, 2025 |
| 21 | GPT-5 mini | OpenAI | 19.4% | 22 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | Aug 7, 2025 |
| 22 | Gemini 2.5 Pro (Mar 2025) | 18.2% | 25 | Compiled by Epoch AI; effort: Mar 2025 | Third-party compilation via Epoch AI | Mar 25, 2025 | |
| 23 | o4-mini | OpenAI | 18.1% | 26 | Compiled by Epoch AI; effort: high | Third-party compilation via Epoch AI | Apr 16, 2025 |
| 24 | Gemini 2.5 Pro (May 2025) | 17.8% | 27 | Compiled by Epoch AI; effort: Jun 2025 | Third-party compilation via Epoch AI | May 6, 2025 | |
| 25 | Claude Sonnet 4.5 | Anthropic | 13.7% | 29 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | Sep 29, 2025 |
| 26 | Gemini 2.5 Flash (Apr 2025) | 12.1% | 30 | Compiled by Epoch AI; effort: Apr 2025 | Third-party compilation via Epoch AI | Apr 17, 2025 | |
| 27 | Claude Opus 4.1 | Anthropic | 11.5% | 31 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | Aug 5, 2025 |
| 28 | Gemini 2.5 Flash (May 2025) | 11.0% | 32 | Compiled by Epoch AI | Third-party compilation via Epoch AI | May 20, 2025 | |
| 29 | Claude Opus 4 | Anthropic | 10.7% | 33 | Compiled by Epoch AI | Third-party compilation via Epoch AI | May 22, 2025 |
| 30 | Gemini 3.1 Flash-Lite | 8.6% | 34 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Mar 3, 2026 | |
| 31 | GLM-4.5OW | Zhipu AI (Z.ai) | 8.3% | 35 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Aug 3, 2025 |
| 32 | GLM-4.5-AirOW | Zhipu AI (Z.ai) | 8.1% | 36 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Jul 20, 2025 |
| 33 | o1-pro | OpenAI | 8.1% | 36 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Mar 19, 2025 |
| 34 | Claude 3.7 Sonnet | Anthropic | 8.0% | 38 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | Feb 24, 2025 |
| 35 | o1 | OpenAI | 8.0% | 39 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | Dec 17, 2024 |
| 36 | Claude Sonnet 4 | Anthropic | 7.8% | 40 | Compiled by Epoch AI; effort: unknown thinking | Third-party compilation via Epoch AI | May 22, 2025 |
| 37 | Gemini 2.0 Flash Thinking (Jan 2025) | 6.6% | 42 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Jan 21, 2025 | |
| 38 | Llama 4 MaverickOW | Meta | 5.7% | 43 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Apr 5, 2025 |
| 39 | GPT-4.5 | OpenAI | 5.4% | 44 | Compiled by Epoch AI; effort: Feb 2025 | Third-party compilation via Epoch AI | Feb 27, 2025 |
| 40 | GPT-4.1 | OpenAI | 5.4% | 45 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Apr 14, 2025 |
| 41 | Gemini 1.5 Pro (Sept 2024) | 4.6% | 46 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Sep 24, 2024 | |
| 42 | Mistral Medium 3 | Mistral AI | 4.5% | 47 | Compiled by Epoch AI | Third-party compilation via Epoch AI | May 7, 2025 |
| 43 | Amazon Nova Pro | Amazon | 4.4% | 48 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Dec 3, 2024 |
| 44 | Claude 3.5 Sonnet (October 2024) | Anthropic | 4.1% | 49 | Compiled by Epoch AI; effort: Oct 2024 | Third-party compilation via Epoch AI | Oct 22, 2024 |
| 45 | Amazon Nova Lite | Amazon | 3.6% | 50 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Dec 3, 2024 |
| 46 | GPT-4o (Nov 2024) | OpenAI | 2.7% | 51 | Compiled by Epoch AI; effort: Nov 2024 | Third-party compilation via Epoch AI | Nov 20, 2024 |
Best score over time
Each step is a new best result in our data, by evaluation date.
- 2024-09-244.6%Gemini 1.5 Pro (Sept 2024)
- 2024-12-178.0%o1
- 2025-02-248.0%Claude 3.7 Sonnet
- 2025-03-198.1%o1-pro
- 2025-03-2518.2%Gemini 2.5 Pro (Mar 2025)
- 2025-04-1620.3%o3
- 2025-06-0521.6%Gemini 2.5 Pro (Jun 2025)
- 2025-08-0725.3%GPT-5
- 2025-10-0731.6%GPT-5 Pro
- 2025-11-1837.5%Gemini 3 Pro
- 2026-02-1946.4%Gemini 3.1 Pro
- 2026-09-0146.5%Claude Fable 5.1
- 2026-09-0354.8%GPT-6 Astra
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 Humanity's Last Exam measure?
About 2,500 expert-written questions across 100+ subjects from the Center for AI Safety and Scale AI, filtered to stump frontier models at creation time. Accuracy (mostly exact-match answers). Data compiled by Epoch AI.
How reliable is Humanity's Last Exam?
Independent audits reported incorrect reference answers in some subjects; settings (tools, text-only) differ between reports. Contamination risk: Low – Private held-out set exists; public questions carry a canary string.