MMLU Pro
General QA · 2024-06-03
Micro-accuracy on the 14-subject, 12,032-question MMLU Pro academic benchmark. Questions have up to ten answer options. Uniform guessing among each question's actual options gives an expected 11.087694718844986 percent, computed from the official TIGER-Lab/MMLU-Pro test parquet at revision b189ec765aa7ed75c8acfea42df31fdae71f97be.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 93.4 |
| Claude Fable 5.1 | 93.0 |
| Claude Opus 5 | 92.0 |
| GPT-6 Astra | 91.5 |
| Gemini 3.1 Pro Preview | 91.3 |
| Claude Sonnet 5.5 | 91.0 |
| GPT-6.1 Sol | 91.0 |
| Gemini 3.8 Flash | 90.7 |
| Gemini 3.7 Flash | 90.6 |
| Opus 4.8 | 90.0 |
| Gemini 3.5 Flash | 90.0 |
| Gemini 3.6 Flash | 90.0 |
| Opus 4.7 | 89.9 |
| GPT-5.6 Sol Pro | 89.9 |
| Muse Spark 1.1 | 89.9 |
| Gemini 3 Pro Preview | 89.8 |
| Opus 4.5 | 89.5 |
| Grok 4.6 | 89.4 |
| Kimi K3 | 89.3 |
| Gemini 3 Flash Preview | 89.0 |
| GPT-5.5 | 88.9 |
| GPT-6 Sol | 88.9 |
| Grok 4.7 | 88.8 |
| Sonnet 5 | 88.5 |
| DeepSeek V4.1 Flash | 88.2 |
| GPT-5.4 | 88.2 |
| Grok 4.5 | 88.2 |
| GPT-5.6 Terra | 88.1 |
| Sonnet 4.6 | 88.0 |
| DeepSeek V4 Pro 0813 | 88.0 |
| GLM-5.3 | 88.0 |
| Qwen3.5 397B A17B | 88.0 |
| Opus 4.1 | 88.0 |
| MiniMax M2.1 | 87.5 |
| Sonnet 4.5 | 87.5 |
| Claude Haiku 5.5 | 87.4 |
| GPT-5.2 | 87.4 |
| Opus 4 | 87.3 |
| GLM-5.3 Flash | 87.3 |
| GPT-5 | 87.1 |