Maths · %
MATH Level 5
The hardest (level 5) problems of the MATH competition dataset (Hendrycks et al., 2021).
Saturated Contamination risk: High 104 results · 94 models Updated Oct 22, 2026 Newest result: 2025-10-30 · stale Official leaderboard ↗
What it measures
Accuracy, Epoch AI's own runs.
Limitations
Near saturation for reasoning models.
Contamination risk: High
Public since 2021.
Leaderboard
94 models
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | GPT-5 | OpenAI | 98.1% | 1 | Epoch AI run, accuracy; effort: high | Independent run | Oct 29, 2025 |
| 2 | GPT-5 mini | OpenAI | 97.8% | 3 | Epoch AI run, accuracy; effort: high | Independent run | Oct 30, 2025 |
| 3 | o4-mini | OpenAI | 97.8% | 4 | Epoch AI run, accuracy; effort: high | Independent run | Apr 16, 2025 |
| 4 | o3 | OpenAI | 97.8% | 5 | Epoch AI run, accuracy; effort: high | Independent run | Apr 16, 2025 |
| 5 | Claude Sonnet 4.5 | Anthropic | 97.7% | 6 | Epoch AI run, accuracy; effort: 32k thinking | Independent run | Oct 21, 2025 |
| 6 | Qwen3-Max | Alibaba (Qwen) | 97.1% | 7 | Epoch AI run, accuracy | Independent run | Oct 9, 2025 |
| 7 | DeepSeek-R1 (May 2025)OW | DeepSeek | 96.6% | 9 | Epoch AI run, accuracy; effort: May 2025 | Independent run | May 29, 2025 |
| 8 | o3-mini | OpenAI | 96.5% | 10 | Epoch AI run, accuracy; effort: high | Independent run | Feb 13, 2025 |
| 9 | Claude Haiku 4.5 | Anthropic | 96.4% | 11 | Epoch AI run, accuracy | Independent run | Oct 22, 2025 |
| 10 | Gemini 2.5 Pro (May 2025) | 95.9% | 12 | Epoch AI run, accuracy; effort: Jun 2025 | Independent run | May 8, 2025 | |
| 11 | Gemini 2.5 Pro (Mar 2025) | 95.6% | 13 | Epoch AI run, accuracy; effort: Mar 2025 | Independent run | May 7, 2025 | |
| 12 | GPT-5 nano | OpenAI | 95.2% | 14 | Epoch AI run, accuracy; effort: medium | Independent run | Aug 20, 2025 |
| 13 | o1 | OpenAI | 94.7% | 17 | Epoch AI run, accuracy; effort: high | Independent run | Feb 13, 2025 |
| 14 | DeepSeek-R1OW | DeepSeek | 93.0% | 19 | Epoch AI run, accuracy | Independent run | Jan 31, 2025 |
| 15 | Claude 3.7 Sonnet | Anthropic | 91.2% | 20 | Epoch AI run, accuracy; effort: 64k thinking | Independent run | Mar 13, 2025 |
| 16 | Grok-3 mini | xAI | 90.9% | 21 | Epoch AI run, accuracy | Independent run | Apr 10, 2025 |
| 17 | DeepSeek-R1-Distill-Llama-70BOW | DeepSeek | 89.9% | 23 | Epoch AI run, accuracy | Independent run | Mar 10, 2025 |
| 18 | o1-mini | OpenAI | 89.2% | 24 | Epoch AI run, accuracy; effort: high | Independent run | Feb 13, 2025 |
| 19 | Grok 3 | xAI | 88.8% | 25 | Epoch AI run, accuracy | Independent run | Apr 10, 2025 |
| 20 | GPT-4.1 mini | OpenAI | 87.3% | 26 | Epoch AI run, accuracy | Independent run | Apr 14, 2025 |
| 21 | DeepSeek-R1-Distill-Qwen-14BOW | DeepSeek | 87.1% | 27 | Epoch AI run, accuracy | Independent run | Mar 11, 2025 |
| 22 | Claude Opus 4 | Anthropic | 85.0% | 29 | Epoch AI run, accuracy | Independent run | May 22, 2025 |
| 23 | Claude Sonnet 4 | Anthropic | 84.4% | 30 | Epoch AI run, accuracy | Independent run | May 22, 2025 |
| 24 | Gemini 2.0 Pro | 83.5% | 32 | Epoch AI run, accuracy; effort: Feb 2025 | Independent run | Feb 6, 2025 | |
| 25 | GPT-4.1 | OpenAI | 83.0% | 33 | Epoch AI run, accuracy | Independent run | Apr 14, 2025 |
| 26 | Gemini 2.0 Flash (Feb 2025) | 82.2% | 34 | Epoch AI run, accuracy; effort: Feb 2025 | Independent run | Feb 6, 2025 | |
| 27 | o1-preview | OpenAI | 81.7% | 35 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 28 | Mistral Medium 3 | Mistral AI | 81.6% | 36 | Epoch AI run, accuracy | Independent run | May 7, 2025 |
| 29 | GPT-4.5 | OpenAI | 78.6% | 37 | Epoch AI run, accuracy; effort: Feb 2025 | Independent run | Feb 28, 2025 |
| 30 | DeepSeek-V3 (Mar 2025)OW | DeepSeek | 75.5% | 38 | Epoch AI run, accuracy; effort: Mar 2025 | Independent run | Apr 1, 2025 |
| 31 | Gemma 3 27BOW | 74.0% | 39 | Epoch AI run, accuracy | Independent run | Mar 13, 2025 | |
| 32 | Llama 4 MaverickOW | Meta | 73.0% | 40 | Epoch AI run, accuracy; effort: FP8 | Independent run | Apr 8, 2025 |
| 33 | Gemini 1.5 Pro (Sept 2024) | 70.4% | 41 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 | |
| 34 | GPT-4.1 nano | OpenAI | 70.0% | 42 | Epoch AI run, accuracy | Independent run | Apr 14, 2025 |
| 35 | Qwen3-235B-A22BOW | Alibaba (Qwen) | 68.9% | 43 | Epoch AI run, accuracy | Independent run | Jun 3, 2025 |
| 36 | Qwen2.5-Max | Alibaba (Qwen) | 67.2% | 45 | Epoch AI run, accuracy | Independent run | Apr 1, 2025 |
| 37 | Qwen Plus (Jan 2025) | Alibaba (Qwen) | 65.3% | 46 | Epoch AI run, accuracy | Independent run | Apr 7, 2025 |
| 38 | Phi-4OW | Microsoft Research | 64.9% | 47 | Epoch AI run, accuracy | Independent run | Jan 31, 2025 |
| 39 | DeepSeek-V3OW | DeepSeek | 64.8% | 48 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 40 | Grok-2 (Dec 2024) | xAI | 63.5% | 49 | Epoch AI run, accuracy | Independent run | Feb 17, 2025 |
| 41 | Qwen2.5-72BOW | Alibaba (Qwen) | 63.2% | 50 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 42 | Llama 4 ScoutOW | Meta | 62.3% | 51 | Epoch AI run, accuracy | Independent run | Apr 8, 2025 |
| 43 | Gemini 1.5 Flash (Sep 2024) | 61.9% | 52 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 | |
| 44 | Claude 3.5 Sonnet (October 2024) | Anthropic | 57.0% | 53 | Epoch AI run, accuracy; effort: Oct 2024 | Independent run | Jan 27, 2025 |
| 45 | Qwen-Turbo | Alibaba (Qwen) | 56.2% | 54 | Epoch AI run, accuracy | Independent run | Apr 7, 2025 |
| 46 | Qwen2.5-32BOW | Alibaba (Qwen) | 56.1% | 55 | Epoch AI run, accuracy | Independent run | Jan 30, 2025 |
| 47 | GPT-4o (Aug 2024) | OpenAI | 53.3% | 56 | Epoch AI run, accuracy; effort: Aug 2024 | Independent run | Jan 27, 2025 |
| 48 | GPT-4o mini | OpenAI | 52.6% | 57 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 49 | Claude 3.5 Sonnet | Anthropic | 51.7% | 58 | Epoch AI run, accuracy; effort: Jun 2024 | Independent run | Jan 27, 2025 |
| 50 | GPT-4o (May 2024) | OpenAI | 51.0% | 59 | Epoch AI run, accuracy; effort: May 2024 | Independent run | Jan 27, 2025 |
| 51 | Mistral Large 2 (Nov 2024)OW | Mistral AI | 50.3% | 60 | Epoch AI run, accuracy | Independent run | Feb 25, 2025 |
| 52 | GPT-4o (Nov 2024) | OpenAI | 49.8% | 61 | Epoch AI run, accuracy; effort: Nov 2024 | Independent run | Feb 5, 2025 |
| 53 | Llama 3.1-405BOW | Meta | 49.8% | 61 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 54 | Mistral Small 3.1OW | Mistral AI | 46.8% | 63 | Epoch AI run, accuracy | Independent run | Mar 18, 2025 |
| 55 | GPT-4 Turbo (Apr 2024) | OpenAI | 46.7% | 64 | Epoch AI run, accuracy | Independent run | Feb 27, 2025 |
| 56 | Claude 3.5 Haiku | Anthropic | 46.4% | 65 | Epoch AI run, accuracy; effort: Oct 2024 | Independent run | Mar 12, 2025 |
| 57 | Mistral Small 3OW | Mistral AI | 44.8% | 66 | Epoch AI run, accuracy | Independent run | Jan 30, 2025 |
| 58 | Mistral Large 2 (Jul 2024)OW | Mistral AI | 44.8% | 66 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 59 | Tulu 3 (Tülu 3) 70BOW | Ai2 | 42.7% | 68 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 60 | Llama 3.3 70BOW | Meta | 41.6% | 69 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 61 | Gemini 1.5 Pro (May 2024) | 40.8% | 70 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 | |
| 62 | GPT-4 Turbo (Nov 2023) | OpenAI | 40.0% | 71 | Epoch AI run, accuracy; effort: Nov 2023 | Independent run | Jan 27, 2025 |
| 63 | Llama 3.2 90BOW | Meta | 39.4% | 72 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 64 | Qwen2-72BOW | Alibaba (Qwen) | 39.1% | 73 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 65 | Claude 3 Opus | Anthropic | 37.5% | 74 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 66 | Llama 3.1-70BOW | Meta | 36.7% | 75 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 67 | Gemma 2 27BOW | 27.9% | 77 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 | |
| 68 | WizardLM-2 8x22BOW | Microsoft | 25.7% | 78 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 69 | Yi-1.5-34BOW | 01.AI | 25.5% | 79 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 70 | Gemini 1.5 Flash (May 2024) | 25.1% | 80 | Epoch AI run, accuracy; effort: May 2024 | Independent run | Jan 27, 2025 | |
| 71 | Mistral Large | Mistral AI | 24.5% | 81 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 72 | Mixtral 8x22BOW | Mistral AI | 24.2% | 82 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 73 | GPT-4 (Jun 2023) | OpenAI | 23.0% | 83 | Epoch AI run, accuracy; effort: Jun 2023 | Independent run | Jan 27, 2025 |
| 74 | Llama 3.1-8BOW | Meta | 22.9% | 84 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 75 | Hermes 2 Theta Llama-3 70BOW | Nous Research | 22.7% | 85 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 76 | Llama 3-70BOW | Meta | 22.6% | 86 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 77 | Gemma 2 9BOW | 21.0% | 87 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 | |
| 78 | Claude 3 Sonnet | Anthropic | 18.2% | 88 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 79 | phi-3-medium 14BOW | Microsoft | 17.6% | 89 | Epoch AI run, accuracy | Independent run | Jan 31, 2025 |
| 80 | GPT-3.5 Turbo (Nov 2023) | OpenAI | 15.9% | 90 | Epoch AI run, accuracy; effort: Nov 2023 | Independent run | Jan 27, 2025 |
| 81 | Ministral 8BOW | Mistral AI | 14.9% | 91 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 82 | Claude 3 Haiku | Anthropic | 14.9% | 92 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 83 | Ministral 3B | Mistral AI | 14.4% | 93 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 84 | Claude 2 | Anthropic | 11.7% | 94 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 85 | DBRXOW | Databricks | 11.7% | 95 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 86 | GPT-3.5 Turbo (Jan 2024) | OpenAI | 11.6% | 96 | Epoch AI run, accuracy; effort: Jan 2024 | Independent run | Jan 27, 2025 |
| 87 | Gemini 1.0 Pro | 11.2% | 97 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 | |
| 88 | Mistral NeMoOW | Mistral AI | 10.8% | 98 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 89 | Mixtral 8x7BOW | Mistral AI | 9.3% | 99 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 90 | DeepSeek LLM 67BOW | DeepSeek | 6.4% | 100 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 91 | Llama 3-8BOW | Meta | 6.1% | 101 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 92 | Yi-34BOW | 01.AI | 5.2% | 102 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 93 | Mistral 7B v0.3OW | Mistral AI | 3.6% | 103 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
| 94 | Llama 2-70BOW | Meta | 3.3% | 104 | Epoch AI run, accuracy | Independent run | Jan 27, 2025 |
Best score over time
Each step is a new best result in our data, by evaluation date.
- 2025-01-2794.4%o1
- 2025-01-3195.2%o3-mini
- 2025-02-1396.5%o3-mini
- 2025-04-1697.8%o4-mini
- 2025-08-2097.9%GPT-5
- 2025-10-2998.1%GPT-5
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 MATH Level 5 measure?
The hardest (level 5) problems of the MATH competition dataset (Hendrycks et al., 2021). Accuracy, Epoch AI's own runs.
How reliable is MATH Level 5?
Near saturation for reasoning models. Contamination risk: High – Public since 2021.