FrontierMath Tier 4
The hardest FrontierMath tier: problems that take expert mathematicians days. Run by Epoch AI.
- Published by
- Epoch AI Benchmarking Hub
- Category
- Maths
- Index weight
- 0.5
- Models
- 59
- Data as of
- 9 Sept 2026
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
- 1GPT-6 Astra97.6%
- 2GPT-6 Astra97.6%
- 3GPT-6 Astra97.6%
- 4GPT-6 Astra97.6%
- 5Claude Fable 590.2%
- 6Claude Fable 5.187.8%
- 7GPT-6 Astra87.8%
- 8GPT-5.6 Sol82.9%
- 9GPT-6 Astra82.9%
- 10GPT-5.6 Sol80.5%
- 11GPT-5.5 Pro78.0%
- 12AI Co-Mathematician75.6%
- 13Claude Opus 573.2%
- 14GPT-5.572.5%
- 15GPT-5.6 Terra70.7%
59 of 59
| # | |||||
|---|---|---|---|---|---|
| 1 |  GPT-6 AstraOpenAI | 97.6% | – | 70.6 | – |
| 2 | GPT-6 AstraOpenAI | 97.6% | – | 69.5 | – |
| 3 | GPT-6 AstraOpenAI | 97.6% | – | 68.6 | – |
| 4 | GPT-6 AstraOpenAI | 97.6% | – | 67.4 | – |
| 5 |  Claude Fable 5Anthropic | 90.2% | – | 65.8 | – |
| 6 | Claude Fable 5.1Anthropic | 87.8% | – | 70.2 | – |
| 7 | GPT-6 AstraOpenAI | 87.8% | – | 65.8 | – |
| 8 | GPT-5.6 SolOpenAI | 82.9% | – | 65.5 | – |
| 9 | GPT-6 AstraOpenAI | 82.9% | – | 64.4 | – |
| 10 | GPT-5.6 SolOpenAI | 80.5% | – | – | – |
| 11 | GPT-5.5 ProOpenAI | 78.0% | – | – | – |
| 12 |  AI Co-MathematicianGoogle | 75.6% | – | – | – |
| 13 | Claude Opus 5Anthropic | 73.2% | – | 66.8 | – |
| 14 | GPT-5.5OpenAI | 72.5% | – | 63.7 | – |
| 15 | GPT-5.6 TerraOpenAI | 70.7% | – | 58.0 | – |
| 16 | GPT-5.6 LunaOpenAI | 61.0% | – | 56.5 | – |
| 17 | GPT-5.4 ProOpenAI | 58.5% | – | 61.1 | – |
| 18 | Claude Opus 4.8Anthropic | 56.1% | – | 59.4 | – |
| 19 | GPT-5.4OpenAI | 49.0% | – | 61.3 | – |
| 20 |  Qwen3 8Alibaba | 46.3% | – | 58.0 | – |
| 21 | GPT-5.2 ProOpenAI | 46.0% | – | – | – |
| 22 |  Kimi K3Moonshot AIopen | 39.0% | – | 62.7 | – |
| 23 | Gemini 3.7 FlashGoogle | 36.6% | – | 60.8 | – |
| 24 | Qwen3 7Alibaba | 34.1% | – | 58.9 | – |
| 25 |  Grok 4.6xAI | 31.7% | – | 62.3 | – |
| 26 | Claude Opus 4.7Anthropic | 31.7% | – | 59.3 | – |
| 27 | GPT-5.2OpenAI | 31.7% | – | 57.3 | – |
| 28 |  GLM-5.3Zhipu AIopen | 29.3% | – | 59.9 | – |
| 29 | GLM-5.2Zhipu AIopen | 29.3% | – | 57.6 | – |
| 30 | Claude Sonnet 5Anthropic | 29.3% | – | 50.7 | – |
| 31 | Gemini 3.1 ProGoogle | 26.8% | – | 63.9 | – |
| 32 | Gemini 3.5 FlashGoogle | 26.8% | – | 60.4 | – |
| 33 |  DeepSeek V4 ProDeepSeekopen | 26.8% | max, 0813 | 56.1 | – |
| 34 | Claude Opus 4.6Anthropic | 26.8% | – | 54.1 | – |
| 35 | Kimi K2.6Moonshot AIopen | 25.6% | kimi-k2.6 | 60.6 | – |
| 36 | DeepSeek V4 FlashDeepSeekopen | 24.4% | max, 0731 | 54.7 | – |
| 37 | Grok 4.5xAI | 24.4% | – | 54.2 | – |
| 38 | Gemini 3.8 FlashGoogle | 21.9% | – | 60.1 | – |
| 39 | Gemini 3.6 FlashGoogle | 21.9% | – | 58.6 | – |
| 40 | GPT-5OpenAI | 21.9% | – | 54.5 | – |
| 41 | GPT-5 ProOpenAI | 19.5% | – | – | – |
| 42 | Gemini 3 FlashGoogle | 17.1% | – | 56.6 | – |
| 43 | GLM-5.3-FlashZhipu AIopen | 17.1% | – | 55.6 | – |
| 44 | Grok 4.20xAI | 17.1% | – | 54.3 | – |
| 45 | Inkling-SmallThinking Machinesopen | 17.1% | – | 48.5 | – |
| 46 | Grok 4.3xAI | 14.6% | – | 50.1 | – |
| 47 | Kimi K2.7 CodeMoonshot AIopen | 12.2% | – | 56.2 | – |
| 48 | GPT-5 miniOpenAI | 12.2% | – | 52.3 | – |
| 49 | GPT-5.4 nanoOpenAI | 12.2% | – | 48.3 | – |
| 50 | GPT-5.4 miniOpenAI | 9.8% | – | 46.0 | – |
| 51 | Claude Opus 4.5Anthropic | 4.9% | – | 58.0 | – |
| 52 | o4-miniOpenAI | 4.9% | – | 48.0 | – |
| 53 | GPT-5.5 InstantOpenAI | 2.4% | – | 60.4 | – |
| 54 | Claude Opus 4.1Anthropic | 2.4% | – | 52.6 | – |
| 55 | Claude Sonnet 4.5Anthropic | 2.4% | – | 52.3 | – |
| 56 | GPT-5 nanoOpenAI | 2.4% | – | 45.9 | – |
| 57 | Gemini 2.5 ProGoogle | 0.0% | – | 54.1 | – |
| 58 | o3-miniOpenAI | 0.0% | – | 47.8 | – |
| 59 | Gemini 3.5 Flash LiteGoogle | 0.0% | – | 44.0 | – |