MATH Level 5
Hardest tier of the MATH dataset. Largely saturated, kept at half weight. Run by Epoch AI.
As of 25 Oct 2026, GPT-5 leads MATH Level 5 on BenchLeader with 98.1%, ahead of GPT-5 at 97.9%, across 96 model configurations with a published result.
- Published by
- Epoch AI Benchmarking Hub
- Category
- Maths
- Index weight
- 0.5
- Models
- 96
- Data as of
- 25 Oct 2026
- Newest result
- 15 Oct 2025
The publisher has posted no new results on this board since 15 Oct 2025. BenchLeader re-reads it every day, so newer models appear here as soon as they are published.
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
What the test looks like
The hardest difficulty tier of the MATH dataset: competition problems across algebra, geometry, number theory, counting and calculus, each with a short final answer.
How it is scored
Percent of problems answered correctly, run by Epoch AI. BenchLeader gives it half weight in the mathematics category.
What to keep in mind
Frontier models score above 95%, so it no longer separates them and mainly identifies weaker or smaller models. The dataset is public and old enough to be in most training sets.
- 1GPT-5 (high)98.1%
- 2GPT-5 (medium)97.9%
- 3GPT-5 mini (high)97.8%
- 4o4-mini (high)97.8%
- 5o3 (high)97.8%
- 6Claude Sonnet 4.597.7%
- 7Qwen3 Max (max)97.1%
- 8GPT-5 mini (medium)96.8%
- 9DeepSeek R1 052896.6%
- 10o3-mini (high)96.5%
- 11Claude Haiku 4.596.4%
- 12Gemini 2.5 Pro95.9%
- 13GPT-5 nano (medium)95.2%
- 14o3-mini (medium)95.2%
- 15GPT-5 nano (high)94.9%
| # | ||||
|---|---|---|---|---|
| 1 |  GPT-5high | 98.1% | 57.9 | 2025-08-07 |
| 2 | GPT-5medium | 97.9% | 57.5 | 2025-08-07 |
| 3 | GPT-5 minihigh | 97.8% | 54.1 | 2025-08-07 |
| 4 | o4-minihigh | 97.8% | 50.1 | 2025-04-16 |
| 5 | o3high | 97.8% | 53.9 | 2025-04-16 |
| 6 |  Claude Sonnet 4.5 | 97.7% | 53.7 | 2025-09-29 |
| 7 |  Qwen3 Maxmax | 97.1% | 52.8 | 2025-09-24 |
| 8 | GPT-5 minimedium | 96.8% | 53.8 | 2025-08-07 |
| 9 |  DeepSeek R1 0528open ↗ | 96.6% | 49.5 | 2025-05-28 |
| 10 | o3-minihigh | 96.5% | 47.2 | 2025-01-31 |
| 11 | Claude Haiku 4.5 | 96.4% | 45.3 | 2025-10-15 |
| 12 |  Gemini 2.5 Pro | 95.9% | 53.9 | 2025-05-06 |
| 13 | GPT-5 nanomedium | 95.2% | 45.9 | 2025-08-07 |
| 14 | o3-minimedium | 95.2% | 45.2 | 2025-01-31 |
| 15 | GPT-5 nanohigh | 94.9% | 47.4 | 2025-08-07 |
| 16 | o1high | 94.7% | 48.3 | 2024-12-17 |
| 17 | o1medium | 94.4% | 48.1 | 2024-12-17 |
| 18 | DeepSeek R1open ↗ | 93.0% | 46.8 | 2025-01-20 |
| 19 | Claude 3.7 Sonnet | 91.2% | 50.0 | 2025-02-24 |
| 20 |  Grok 3 minilow | 90.9% | 49.8 | 2025-04-09 |
| 21 | R1 Distill Llama 70Bopen ↗ | 89.9% | 41.5 | 2025-01-20 |
| 22 | o1-minihigh | 89.2% | – | 2024-09-12 |
| 23 | Grok 3 | 88.8% | 51.5 | 2025-04-09 |
| 24 | Grok 3 minihigh | 88.1% | 52.6 | 2025-04-09 |
| 25 | GPT-4.1 mini | 87.3% | 45.0 | 2025-04-14 |
| 26 | DeepSeek-R1-Distill-Qwen-14Bopen ↗ | 87.1% | 41.2 | 2025-01-20 |
| 27 | Claude Opus 4 | 85.0% | 52.6 | 2025-05-22 |
| 28 | Claude Sonnet 4 | 84.4% | 50.5 | 2025-05-22 |
| 29 | o1-minimedium | 84.3% | 44.4 | 2024-09-12 |
| 30 | Gemini 2.0 Pro | 83.5% | 47.0 | 2025-02-05 |
| 31 | GPT-4.1 | 83.0% | 48.3 | 2025-04-14 |
| 32 | Gemini 2.0 Flash | 82.2% | 43.2 | 2025-02-05 |
| 33 | o1 | 81.7% | 53.1 | 2024-09-12 |
| 34 |  Mistral Medium 3 | 81.6% | 44.1 | 2025-05-07 |
| 35 | GPT-4.5 | 78.6% | 49.7 | 2025-02-27 |
| 36 | Gemma 3 27Bopen ↗ | 74.0% | 38.9 | 2025-03-12 |
| 37 |  Llama 4 Maverickopen ↗ | 73.0% | 42.8 | 2025-04-05 |
| 38 | Gemini 1.5 Pro 002 | 70.4% | 41.6 | 2024-09-24 |
| 39 | GPT-4.1 nano | 70.0% | 38.0 | 2025-04-14 |
| 40 | Qwen3 235B A22Bopen ↗ | 68.9% | 50.1 | 2025-04-29 |
| 41 | Qwen-Plus | 65.3% | 48.5 | 2025-01-25 |
| 42 |  Phi-4open ↗ | 64.9% | 38.9 | 2024-12-12 |
| 43 | DeepSeek V3open ↗ | 64.8% | 44.0 | 2024-12-26 |
| 44 | Grok 2open | 63.5% | 41.9 | 2024-12-12 |
| 45 | Qwen2.5 72Bopen ↗ | 63.2% | 42.2 | 2024-09-19 |
| 46 | Llama 4 Scoutopen ↗ | 62.3% | 39.7 | 2025-04-05 |
| 47 | Gemini 1.5 Flash 002 | 61.9% | 38.3 | 2024-09-24 |
| 48 | Claude 3.5 Sonnet | 57.0% | 46.3 | 2024-10-22 |
| 49 | Qwen Turbo | 56.2% | – | 2024-11-01 |
| 50 | Qwen2.5 32B Instructopen ↗ | 56.1% | – | 2024-09-17 |
| 51 | GPT-4o | 53.3% | 43.9 | 2024-08-06 |
| 52 | GPT-4o mini | 52.6% | 35.8 | 2024-07-18 |
| 53 | Mistral Large 2.1open | 50.3% | 42.5 | 2024-11-18 |
| 54 | Llama 3.1 405Bopen ↗ | 49.8% | 42.9 | 2024-07-23 |
| 55 | Mistral Small 3.1open | 46.8% | 36.3 | 2025-03-17 |
| 56 | GPT-4 Turbo | 46.7% | 40.9 | 2024-04-09 |
| 57 | Claude 3.5 Haiku | 46.4% | 39.7 | 2024-10-22 |
| 58 | Mistral Large 2open | 44.8% | 40.7 | 2024-07-24 |
| 59 | Mistral Small 3open ↗ | 44.8% | 39.3 | 2025-01-25 |
| 60 |  Llama 3.1 Tulu 3 70B Dpo | 42.7% | – | 2024-11-21 |
More maths benchmarks · All LLM benchmarks
Cite as: BenchLeader, “MATH Level 5 leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/math_level_5.
MATH Level 5: questions
- What does MATH Level 5 measure?
- The hardest difficulty tier of the MATH dataset: competition problems across algebra, geometry, number theory, counting and calculus, each with a short final answer. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads MATH Level 5?
- GPT-5 leads MATH Level 5 with 98.1% as of 25 Oct 2026, ahead of GPT-5 at 97.9%.
- How many models have MATH Level 5 results?
- 96 model configurations have a MATH Level 5 result on BenchLeader, all taken from Epoch AI Benchmarking Hub.
- Who runs MATH Level 5 and how often is it updated?
- MATH Level 5 is published by Epoch AI Benchmarking Hub. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does MATH Level 5 count toward the BenchLeader Index?
- Yes. MATH Level 5 contributes to the maths category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.