Back to Benchmarks
📐
MATH-500
MATH 500 Subset
500 curated competition-level math problems across 6 domains
📐 Math
Best: 99%
21 Models Evaluated
Champion
Step-3.5-Flash
Stepfun
99.8%
Best Score
Model Rankings
All models ranked by score on MATH-500. Scores are relative to the best score of 99%.
| # | Model | Provider | Score | vs Best |
|---|---|---|---|---|
| 1 | Step-3.5-Flash Elo 1389 | Stepfun | 99.8% | Best |
| 2 | GPT-5.5 Elo 1478 | OpenAI | 99% | +0.0% |
| 3 | GPT-5.4 Elo 1463 | OpenAI | 99% | +0.0% |
| 4 | GPT-5.4 Pro Elo 1480 | OpenAI | 99% | +0.0% |
| 5 | Claude Opus 4.7 Elo 1500 | Anthropic | 98% | -1.0% |
| 6 | Kimi K2.5 Elo 1438 | Moonshot | 98% | -1.0% |
| 7 | GPT-oss 120B Elo 1355 | OpenAI | 97.9% | -1.1% |
| 8 | Claude Sonnet 4.6 Elo 1460 | Anthropic | 97.8% | -1.2% |
| 9 | Claude Opus 4.6 Elo 1503 | Anthropic | 97.6% | -1.4% |
| 10 | DeepSeek R1 Elo 1398 | DeepSeek | 97.3% | -1.7% |
| 11 | DeepSeek V4 Pro (Max) | DeepSeek | 97% | -2.0% |
| 12 | Nemotron Ultra 253B Elo 1348 | NVIDIA | 97% | -2.0% |
| 13 | MiMo-V2-Flash Elo 1393 | Xiaomi | 94.1% | -4.9% |
| 14 | Gemini 3.1 Pro Elo 1492 | 94% | -5.0% | |
| 15 | DeepSeek V3.2 Elo 1423 | DeepSeek | 94% | -5.0% |
| 16 | Grok 3 Elo 1412 | xAI | 93.6% | -5.4% |
| 17 | Mistral Large Elo 1416 | Mistral | 93.6% | -5.4% |
| 18 | Gemini 3 Pro Elo 1486 | 91% | -8.0% | |
| 19 | Qwen 3.5 Elo 1450 | Alibaba | 88% | -11.0% |
| 20 | GLM-5 Elo 1454 | Zhipu AI | 88% | -11.0% |
| 21 | MiniMax M2.5 Elo 1404 | MiniMax | 85% | -14.0% |
Score Distribution
Step-3.5-Flash
99.8%
GPT-5.5
99%
GPT-5.4
99%
GPT-5.4 Pro
99%
Claude Opus 4.7
98%
Kimi K2.5
98%
GPT-oss 120B
97.9%
Claude Sonnet 4.6
97.8%
Claude Opus 4.6
97.6%
DeepSeek R1
97.3%
Run This Benchmark
Learn how to evaluate your own models on MATH-500.
MATH-500 is a standardized benchmark for evaluating AI model capabilities.
To run this benchmark on your own models, visit the official benchmark page or leaderboard for detailed instructions, code, and evaluation scripts.
Data compiled from official model cards, public leaderboards, and independent evaluations.