MATH Level 5
Hardest tier of the MATH dataset. Largely saturated, kept at half weight. Run by Epoch AI.
- Published by
- Epoch AI Benchmarking Hub
- Category
- Maths
- Index weight
- 0.5
- Models
- 90
- 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-598.1%
- 2GPT-597.9%
- 3GPT-5 mini97.8%
- 4o4-mini97.8%
- 5o397.8%
- 6Claude Sonnet 4.597.7%
- 7Qwen3 Max97.1%
- 8GPT-5 mini96.8%
- 9DeepSeek R196.6%
- 10o3-mini96.5%
- 11Claude Haiku 4.596.4%
- 12Gemini 2.5 Pro95.9%
- 13GPT-5 nano95.2%
- 14o3-mini95.2%
- 15GPT-5 nano94.9%
90 of 90
| # | |||||
|---|---|---|---|---|---|
| 1 |  GPT-5OpenAI | 98.1% | – | 54.5 | – |
| 2 | GPT-5OpenAI | 97.9% | – | 58.5 | – |
| 3 | GPT-5 miniOpenAI | 97.8% | – | 52.3 | – |
| 4 | o4-miniOpenAI | 97.8% | – | 48.0 | – |
| 5 | o3OpenAI | 97.8% | – | 52.9 | – |
| 6 |  Claude Sonnet 4.5Anthropic | 97.7% | – | 52.3 | – |
| 7 |  Qwen3 MaxAlibaba | 97.1% | – | 52.7 | – |
| 8 | GPT-5 miniOpenAI | 96.8% | – | 54.5 | – |
| 9 |  DeepSeek R1DeepSeekopen | 96.6% | 0528 | 48.9 | – |
| 10 | o3-miniOpenAI | 96.5% | – | 47.8 | – |
| 11 | Claude Haiku 4.5Anthropic | 96.4% | – | 47.9 | – |
| 12 |  Gemini 2.5 ProGoogle | 95.9% | – | 54.1 | – |
| 13 | GPT-5 nanoOpenAI | 95.2% | – | 46.5 | – |
| 14 | o3-miniOpenAI | 95.2% | – | 45.6 | – |
| 15 | GPT-5 nanoOpenAI | 94.9% | – | 45.9 | – |
| 16 | o1OpenAI | 94.7% | – | 44.3 | – |
| 17 | o1OpenAI | 94.4% | – | 48.3 | – |
| 18 | Claude 3.7 SonnetAnthropic | 91.2% | – | 49.8 | – |
| 19 |  Grok 3 minixAI | 90.9% | – | 50.2 | – |
| 20 | DeepSeek-R1-Distill-Llama-70BDeepSeekopen | 89.9% | – | 42.2 | – |
| 21 | o1-miniOpenAI | 89.2% | – | – | – |
| 22 | Grok 3xAI | 88.8% | – | 50.0 | – |
| 23 | Grok 3 minixAI | 88.1% | – | 53.3 | – |
| 24 | GPT-4.1 miniOpenAI | 87.3% | – | 44.5 | – |
| 25 | DeepSeek-R1-Distill-Qwen-14BDeepSeekopen | 87.1% | – | 41.4 | – |
| 26 | Claude Opus 4Anthropic | 85.0% | – | 50.7 | – |
| 27 | Claude Sonnet 4Anthropic | 84.4% | – | 49.5 | – |
| 28 | o1-miniOpenAI | 84.3% | – | 44.4 | – |
| 29 | Gemini 2.0 ProGoogle | 83.5% | – | 47.0 | – |
| 30 | GPT-4.1OpenAI | 83.0% | – | 47.3 | – |
| 31 | Gemini 2.0 FlashGoogle | 82.2% | – | 42.3 | – |
| 32 | o1OpenAI | 81.7% | – | 53.2 | – |
| 33 |  Mistral Medium (latest)Mistral AI | 81.6% | medium, 2505 | 45.8 | – |
| 34 | GPT-4.5OpenAI | 78.6% | – | 49.6 | – |
| 35 | Gemma 3 27BGoogleopen | 74.0% | – | 39.7 | – |
| 36 |  Llama 4 MaverickMetaopen | 73.0% | – | 42.6 | – |
| 37 | Gemini 1.5 Pro 002Google | 70.4% | – | 42.8 | – |
| 38 | GPT-4.1 nanoOpenAI | 70.0% | – | 38.1 | – |
| 39 | Qwen3 235B A22BAlibabaopen | 68.9% | – | 48.7 | – |
| 40 | Qwen PlusAlibaba | 65.3% | – | 48.4 | – |
| 41 |  Phi-4Microsoftopen | 64.9% | phi-4 | 39.7 | – |
| 42 | DeepSeek V3DeepSeekopen | 64.8% | DeepSeek-V3 | 43.8 | – |
| 43 | Grok 2xAI | 63.5% | 1212 | 41.6 | – |
| 44 | Qwen2.5 72BAlibabaopen | 63.2% | – | 42.1 | – |
| 45 | Llama 4 ScoutMetaopen | 62.3% | – | 39.7 | – |
| 46 | Gemini 1.5 Flash 002Google | 61.9% | – | 38.7 | – |
| 47 | Claude 3.5 SonnetAnthropic | 57.0% | – | 45.1 | – |
| 48 | Qwen TurboAlibaba | 56.2% | – | – | – |
| 49 | Qwen2.5 32B InstructAlibabaopen | 56.1% | – | – | – |
| 50 | GPT-4oOpenAI | 53.3% | – | 42.3 | – |
| 51 | GPT-4o miniOpenAI | 52.6% | – | 35.8 | – |
| 52 | Mistral Large 2Mistral AIopen | 50.3% | 2411 | 41.2 | – |
| 53 | Llama 3.1 405BMetaopen | 49.8% | – | 42.6 | – |
| 54 | Mistral Small (latest)Mistral AI | 46.8% | 2503 | 41.1 | – |
| 55 | GPT-4 TurboOpenAI | 46.7% | – | 41.0 | – |
| 56 | Claude 3.5 HaikuAnthropic | 46.4% | – | 39.8 | – |
| 57 | Llama 3.1 Tulu 3 70B DpoAllen Institute for AI | 42.7% | – | – | – |
| 58 | Llama 3.3 70BMetaopen | 41.6% | – | 40.7 | – |
| 59 | Gemini 1.5 Pro 001Google | 40.8% | – | 44.3 | – |
| 60 | GPT-4OpenAI | 40.0% | preview, 1106 | 41.8 | – |