FrontierMath Tiers 1–3
Unpublished research-level maths problems, tiers 1–3 (v2 private set). Run by Epoch AI.
- Published by
- Epoch AI Benchmarking Hub
- Category
- Maths
- Index weight
- 1.0
- Models
- 103
- 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 Astra93.7%
- 2Claude Fable 5.190.2%
- 3GPT-5.6 Sol89.1%
- 4GPT-5.5 Pro87.7%
- 5Claude Fable 587.0%
- 6GPT-5.6 Terra86.0%
- 7Claude Opus 585.6%
- 8GPT-5.585.3%
- 9GPT-5.4 Pro82.5%
- 10GPT-5.6 Luna82.1%
- 11Claude Opus 4.880.0%
- 12GPT-5.478.6%
- 13Qwen3 874.7%
- 14GPT-5.2 Pro74.0%
- 15Kimi K372.2%
103 of 103
| # | |||||
|---|---|---|---|---|---|
| 1 |  GPT-6 AstraOpenAI | 93.7% | – | 70.6 | – |
| 2 |  Claude Fable 5.1Anthropic | 90.2% | – | 70.2 | – |
| 3 | GPT-5.6 SolOpenAI | 89.1% | – | 65.5 | – |
| 4 | GPT-5.5 ProOpenAI | 87.7% | – | – | – |
| 5 | Claude Fable 5Anthropic | 87.0% | – | 65.8 | – |
| 6 | GPT-5.6 TerraOpenAI | 86.0% | – | 58.0 | – |
| 7 | Claude Opus 5Anthropic | 85.6% | – | 66.8 | – |
| 8 | GPT-5.5OpenAI | 85.3% | – | 63.7 | – |
| 9 | GPT-5.4 ProOpenAI | 82.5% | – | 61.1 | – |
| 10 | GPT-5.6 LunaOpenAI | 82.1% | – | 56.5 | – |
| 11 | Claude Opus 4.8Anthropic | 80.0% | – | 59.4 | – |
| 12 | GPT-5.4OpenAI | 78.6% | – | 61.3 | – |
| 13 |  Qwen3 8Alibaba | 74.7% | – | 58.0 | – |
| 14 | GPT-5.2 ProOpenAI | 74.0% | – | – | – |
| 15 |  Kimi K3Moonshot AIopen | 72.2% | – | 62.7 | – |
| 16 |  Gemini 3.7 FlashGoogle | 71.6% | – | 60.8 | – |
| 17 | Claude Opus 4.7Anthropic | 70.2% | – | 59.3 | – |
| 18 |  GLM-5.3Zhipu AIopen | 68.8% | – | 59.9 | – |
| 19 | Gemini 3.8 FlashGoogle | 68.4% | – | 60.1 | – |
| 20 | GPT-5.2OpenAI | 67.4% | – | 57.3 | – |
| 21 |  Grok 4.6xAI | 66.0% | – | 62.3 | – |
| 22 | Claude Opus 4.6Anthropic | 66.0% | – | 54.1 | – |
| 23 | Claude Sonnet 5Anthropic | 65.6% | – | 50.7 | – |
| 24 | Qwen3 7Alibaba | 64.6% | – | 58.9 | – |
| 25 |  DeepSeek V4 ProDeepSeekopen | 64.6% | max, 0813 | 56.1 | – |
| 26 | Gemini 3.5 FlashGoogle | 62.8% | – | 60.4 | – |
| 27 | Gemini 3.1 ProGoogle | 59.6% | – | 63.9 | – |
| 28 | GLM-5.2Zhipu AIopen | 59.2% | – | 57.6 | – |
| 29 | Gemini 3.6 FlashGoogle | 59.0% | – | 58.6 | – |
| 30 | DeepSeek V4 FlashDeepSeekopen | 57.5% | max, 0731 | 54.7 | – |
| 31 | Kimi K2.6Moonshot AIopen | 57.2% | kimi-k2.6 | 60.6 | – |
| 32 | Grok 4.5xAI | 57.2% | – | 54.2 | – |
| 33 | GLM-5.3-FlashZhipu AIopen | 55.8% | – | 55.6 | – |
| 34 | GPT-5 ProOpenAI | 55.8% | – | – | – |
| 35 | GPT-5OpenAI | 55.4% | – | 54.5 | – |
| 36 | GLM-5.2Zhipu AIopen | 54.7% | – | – | – |
| 37 | Kimi K2.7 CodeMoonshot AIopen | 54.0% | – | 56.2 | – |
| 38 | Gemini 3 FlashGoogle | 51.2% | – | 56.6 | – |
| 39 | GPT-5.4 miniOpenAI | 51.2% | – | 46.0 | – |
| 40 | GPT-5 miniOpenAI | 46.7% | – | 52.3 | – |
| 41 | Inkling-SmallThinking Machinesopen | 46.3% | – | 48.5 | – |
| 42 | Grok 4.20xAI | 44.9% | – | 54.3 | – |
| 43 | GPT-5.4 nanoOpenAI | 44.9% | – | 48.3 | – |
| 44 | Grok 4.3xAI | 42.8% | – | 50.1 | – |
| 45 | GLM-5.2Zhipu AIopen | 42.5% | – | 50.3 | – |
| 46 | GPT-5.6 LunaOpenAI | 41.4% | – | 53.6 | – |
| 47 | GPT-5.6 LunaOpenAI | 39.6% | – | 46.9 | – |
| 48 | Qwen3.6 PlusAlibabaopen | 38.3% | – | 59.0 | – |
| 49 | GPT-5OpenAI | 37.2% | – | 53.1 | – |
| 50 | GLM-5.1Zhipu AIopen | 36.8% | glm-5.1 | 60.1 | – |
| 51 | o4-miniOpenAI | 36.1% | – | 48.0 | – |
| 52 | Qwen3.6 27BAlibabaopen | 35.1% | – | 53.2 | – |
| 53 | Claude Opus 4.5Anthropic | 34.4% | – | 58.0 | – |
| 54 | Qwen3.7 PlusAlibaba | 34.4% | – | – | – |
| 55 | Qwen3.6 27BAlibabaopen | 34.0% | – | 50.7 | – |
| 56 | o3OpenAI | 33.3% | – | 52.9 | – |
| 57 | Qwen3.6 PlusAlibabaopen | 32.3% | – | – | – |
| 58 | Qwen3.5 397B-A17BAlibabaopen | 31.2% | – | 52.5 | – |
| 59 | o3OpenAI | 29.8% | – | 53.6 | – |
| 60 | Qwen3.5 397B-A17BAlibabaopen | 29.5% | – | 58.6 | – |