FrontierMath-Tier-4-2025-07-01-Private
Re-run by Epoch AIThe hardest tier of FrontierMath: unpublished research-level mathematics problems written for Epoch AI.
- Kind
- Run independently by Epoch AI, with public logs. Our own re-runs are not built yet.
- Score shown
- Epoch's column "Best score (across scorers)", as a percentage, higher is better; random guessing scores about 0.0%. Error bars as the source gives them.
- Results
- 72 in all: 9 for open-weight models, 59 for closed ones, 4 where Epoch does not say. 9 linked to a model page here.
Best open-weight result: moonshotai/Kimi-K2.6, 14.6%, number 22 of 72 results.
| # | Model | Weights | Setting | Score | Date | Source |
|---|---|---|---|---|---|---|
| 1 | gdm-ai-co-mathematicianGoogle DeepMind | closed | default | 47.9%±7.2 | 8 May 2026 | Epoch AI |
| 2 | gpt-5.5-pro-pre-release_highOpenAI | closed | effort high | 39.6%±7.1 | 23 Apr 2026 | Epoch AI |
| 3 | gpt-5.5-pro-pre-release_xhighOpenAI | closed | effort xhigh | 39.6%±7.1 | 23 Apr 2026 | Epoch AI |
| 4 | gpt-5.4-pro-2026-03-05-web-appOpenAI | closed | default | 37.5%±7.0 | 6 Mar 2026 | Epoch AI |
| 5 | gpt-5.5-pre-release_xhighOpenAI | closed | effort xhigh | 35.4%±6.9 | 23 Apr 2026 | Epoch AI |
| 6 | gpt-5.2-pro-2025-12-11-webappOpenAI | closed | default | 31.3%±6.7 | 24 Dec 2025 | Epoch AI |
| 7 | claude-opus-4-8_maxAnthropic | closed | effort max | 31.3%±6.8 | 8 Jun 2026 | Epoch AI log |
| 8 | gpt-5.4-2026-03-05_xhighOpenAI | closed | effort xhigh | 27.1%±6.4 | 6 Mar 2026 | Epoch AI |
| 9 | claude-opus-4-7_xhighAnthropic | closed | effort xhigh | 22.9%±6.1 | 17 Apr 2026 | Epoch AI |
| 10 | claude-opus-4-6_maxAnthropic | closed | effort max | 22.9%±6.1 | 12 Feb 2026 | Epoch AI log |
| 11 | claude-opus-4-6_32KAnthropic | closed | default | 20.8%±5.9 | 6 Feb 2026 | Epoch AI log |
| 12 | claude-opus-4-6_64KAnthropic | closed | default | 20.8%±5.9 | 6 Feb 2026 | Epoch AI log |
| 13 | gpt-5.2-2025-12-11_highOpenAI | closed | effort high | 18.8%±5.6 | 11 Dec 2025 | Epoch AI log |
| 14 | gpt-5.2-2025-12-11_xhighOpenAI | closed | effort xhigh | 18.8%±5.6 | 14 Dec 2025 | Epoch AI |
| 15 | gemini-3-pro-previewGoogle DeepMind | closed | default | 18.8%±5.7 | 21 Nov 2025 | Epoch AI |
| 16 | gemini-3.1-pro-previewGoogle DeepMind | closed | default | 16.7%±5.4 | 19 Feb 2026 | Epoch AI log |
| 17 | gpt-5.2-2025-12-11_mediumOpenAI | closed | effort medium | 16.7%±5.4 | 11 Dec 2025 | Epoch AI log |
| 18 | gpt-5-pro-2025-10-06_highOpenAI | closed | effort high | 14.6%±5.1 | 9 Oct 2025 | Epoch AI |
| 19 | muse-sparkMeta AI | closed | default | 14.6%±5.1 | 8 Apr 2026 | Epoch AI |
| 20 | claude-opus-4-6Anthropic | closed | default | 14.6%±5.1 | 5 Feb 2026 | Epoch AI log |
| 21 | gemini-3.5-flash_highGoogle DeepMind | closed | effort high | 14.6%±5.1 | 25 May 2026 | Epoch AI |
| 22 | moonshotai/Kimi-K2.6Epoch: kimi-k2.6, Moonshot | open | default | 14.6%±5.1 | 8 May 2026 | Epoch AI |
| 23 | zai-org/GLM-5.1Epoch: glm-5.1, Z.ai (Zhipu AI) | open | default | 12.5%±4.8 | 12 May 2026 | Epoch AI |
| 24 | gpt-5-2025-08-07_highOpenAI | closed | effort high | 12.5%±4.8 | 30 Oct 2025 | Epoch AI log |
| 25 | gpt-5.1-2025-11-13_highOpenAI | closed | effort high | 12.5%±4.8 | 13 Nov 2025 | Epoch AI log |
| 26 | gemini-2.5-deep-think-2025-08-01-webappGoogle,Google DeepMind | closed | default | 10.4%±4.4 | not dated | Epoch AI |
| 27 | qwen3.6-plusAlibaba | closed | default | 8.3%±4.0 | 12 May 2026 | Epoch AI |
| 28 | claude-sonnet-4-6_16KAnthropic | closed | default | 8.3%±4.0 | 21 Feb 2026 | Epoch AI log |
| 29 | gpt-5-2025-08-07_mediumOpenAI | closed | effort medium | 6.3%±3.5 | 7 Aug 2025 | Epoch AI log |
| 30 | gpt-5-mini-2025-08-07_highOpenAI | closed | effort high | 6.3%±3.5 | 30 Oct 2025 | Epoch AI log |
| 31 | gpt-5.2-2025-12-11_lowOpenAI | closed | effort low | 6.3%±3.5 | 11 Dec 2025 | Epoch AI log |
| 32 | gpt-5.4-nano-2026-03-17_highOpenAI | closed | effort high | 6.3%±3.5 | 15 Apr 2026 | Epoch AI |
| 33 | o4-mini-2025-04-16_highOpenAI | closed | effort high | 6.3%±3.5 | 1 Jul 2025 | Epoch AI log |
| 34 | moonshotai/Kimi-K2.5Epoch: fireworks/kimi-k2p5, Moonshot | open | default | 4.2%±2.9 | 2 Feb 2026 | Epoch AI |
| 35 | claude-opus-4-1-20250805_27KAnthropic | closed | default | 4.2%±2.9 | 5 Aug 2025 | Epoch AI log |
| 36 | claude-opus-4-20250514_27KAnthropic | closed | default | 4.2%±2.9 | 1 Jul 2025 | Epoch AI log |
| 37 | claude-opus-4-5-20251101Anthropic | closed | default | 4.2%±2.9 | 25 Nov 2025 | Epoch AI log |
| 38 | claude-opus-4-5-20251101_32KAnthropic | closed | default | 4.2%±2.9 | 25 Nov 2025 | Epoch AI log |
| 39 | claude-sonnet-4-5-20250929_32KAnthropic | closed | default | 4.2%±2.9 | 22 Oct 2025 | Epoch AI log |
| 40 | gemini-2.5-flashGoogle DeepMind | closed | default | 4.2%±2.9 | 18 Dec 2025 | Epoch AI log |
| 41 | gemini-2.5-proGoogle DeepMind | closed | default | 4.2%±2.9 | 3 Jul 2025 | Epoch AI log |
| 42 | gemini-3-flash-previewGoogle DeepMind | closed | default | 4.2%±2.9 | 17 Dec 2025 | Epoch AI log |
| 43 | gpt-5-mini-2025-08-07_mediumOpenAI | closed | effort medium | 4.2%±2.9 | 7 Aug 2025 | Epoch AI log |
| 44 | gpt-5.1-2025-11-13_mediumOpenAI | closed | effort medium | 4.2%±2.9 | 17 Nov 2025 | Epoch AI log |
| 45 | o3-mini-2025-01-31_highOpenAI | closed | effort high | 4.2%±2.9 | 1 Jul 2025 | Epoch AI log |
| 46 | qwen3.6-max-previewAlibaba | not stated | default | 4.2%±2.9 | 28 May 2026 | Epoch AI |
| 47 | zai-org/GLM-4.6Epoch: zai-org/GLM-4.6, Z.ai (Zhipu AI),Tsinghua University | open | default | 2.1%±2.1 | 8 Dec 2025 | Epoch AI log |
| 48 | deepseek-ai/DeepSeek-V3.2Epoch: fireworks/deepseek-v3p2, DeepSeek | open | default | 2.1%±2.1 | 16 Dec 2025 | Epoch AI |
| 49 | zai-org/GLM-5Epoch: glm-5, Z.ai (Zhipu AI) | open | default | 2.1%±2.1 | 19 Feb 2026 | Epoch AI log |
| 50 | claude-haiku-4-5-20251001_32KAnthropic | closed | default | 2.1%±2.1 | 22 Oct 2025 | Epoch AI log |
| 51 | claude-opus-4-5-20251101_16KAnthropic | closed | default | 2.1%±2.1 | 25 Nov 2025 | Epoch AI log |
| 52 | claude-sonnet-4-5-20250929Anthropic | closed | default | 2.1%±2.1 | 29 Sep 2025 | Epoch AI log |
| 53 | gemini-2.5-pro-preview-06-05Google DeepMind | closed | default | 2.1%±2.1 | 3 Jul 2025 | Epoch AI log |
| 54 | gpt-5-nano-2025-08-07_mediumOpenAI | closed | effort medium | 2.1%±2.1 | 7 Aug 2025 | Epoch AI log |
| 55 | grok-4-0709xAI | closed | default | 2.1%±2.1 | 11 Aug 2025 | Epoch AI log |
| 56 | o3-2025-04-16_highOpenAI | closed | effort high | 2.1%±2.1 | 1 Jul 2025 | Epoch AI log |
| 57 | o4-mini-2025-04-16_mediumOpenAI | closed | effort medium | 2.1%±2.1 | 7 Aug 2025 | Epoch AI log |
| 58 | qwen3.5-plusAlibaba | not stated | default | 2.1%±2.1 | 15 May 2026 | Epoch AI |
| 59 | gpt-5.4-mini-2026-03-17_highOpenAI | closed | effort high | 2.1%±2.1 | 15 Apr 2026 | Epoch AI |
| 60 | grok-4-heavy-web-appxAI | closed | default | 2.1%±2.1 | 9 Oct 2025 | Epoch AI |
| 61 | claude-3-5-sonnet-20240620Anthropic | closed | default | 0.0% | 1 Jul 2025 | Epoch AI log |
| 62 | claude-3-5-sonnet-20241022Anthropic | closed | default | 0.0% | 1 Jul 2025 | Epoch AI log |
| 63 | claude-opus-4-20250514Anthropic | closed | default | 0.0% | 1 Jul 2025 | Epoch AI log |
| 64 | claude-sonnet-4-20250514Anthropic | closed | default | 0.0% | 1 Jul 2025 | Epoch AI log |
| 65 | gpt-4.1-2025-04-14OpenAI | closed | default | 0.0% | 1 Jul 2025 | Epoch AI log |
| 66 | gpt-5-nano-2025-08-07_highOpenAI | closed | effort high | 0.0% | 30 Oct 2025 | Epoch AI log |
| 67 | grok-3-betaxAI | closed | default | 0.0% | 1 Jul 2025 | Epoch AI log |
| 68 | moonshotai/Kimi-K2-ThinkingEpoch: moonshotai/Kimi-K2-Thinking, Moonshot | open | default | 0.0% | 4 Dec 2025 | Epoch AI log |
| 69 | Qwen/Qwen3-235B-A22B-Thinking-2507Epoch: Qwen/Qwen3-235B-A22B-Thinking-2507, Alibaba | open | default | 0.0% | 11 Dec 2025 | Epoch AI log |
| 70 | qwen3.5-flashAlibaba | not stated | default | 0.0% | 12 May 2026 | Epoch AI |
| 71 | qwen3.6-flashAlibaba | not stated | default | 0.0% | 12 May 2026 | Epoch AI |
| 72 | zai-org/GLM-4.7Epoch: zai-org/GLM-4.7, Z.ai (Zhipu AI) | open | default | 0.0% | 30 Jan 2026 | Epoch AI log |
Open-weight rows are shaded. Dates are when the result was run or evaluated, where the data says. Data from Epoch AI (CC-BY 4.0), fetched 19 Nov 2026: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource]. Epoch AI. Method