Benchmark Leaderboard
Top models by Massive Multitask Language Understanding score
Sorts are only valid inside this benchmark. A score gap of this benchmark does not map to the same gap on another benchmark, so validate any short-list by checking peer rows and model context/price fit.
| # | Model | Score | Version | Source |
|---|---|---|---|---|
| 1 | Gemini 3.1 Pro Preview | 98 | MMLU (accuracy) | https://automatio.ai/models/gemini-3-1-pro |
| 2 | GPT-5.5 | 92.4 | MMLU (accuracy) | https://tokenmix.ai/blog/gpt-5-5-spud-review-88-swe-bench-2026 |
| 3 | Claude Opus 4.6 | 91.1 | MMLU (accuracy) | https://automatio.ai/models/claude-opus-4-6 |
| 4 | DeepSeek V4 Pro | 90.1 | 5-shot | https://api-docs.deepseek.com/news/news260424 |
| 5 | Xiaomi MiMo-V2.5-Pro | 89.4 | — | https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro |
| 6 | Claude Sonnet 4.6 | 89.3 | MMLU (accuracy) | https://automatio.ai/models/claude-sonnet-4-6 |
| 7 | Claude 3 Opus | 88.7 | 5-shot | https://crfm.stanford.edu/helm/classic/latest/ |
| 8 | Claude 3.5 Sonnet | 88.7 | 5-shot | https://crfm.stanford.edu/helm/classic/latest/ |
| 9 | GPT-4o (05-13) | 88.7 | 5-shot | https://crfm.stanford.edu/helm/classic/latest/ |
| 10 | DeepSeek V4 Flash | 88.7 | MMLU EM, DeepSeek-V4-Flash-Base, 5-shot | https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash |
| 11 | Llama 3.1 405B | 88.6 | 5-shot | https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard |
| 12 | Llama 3.1 405B Instruct | 88.6 | 5-shot | https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard |
| 13 | DeepSeek V3 | 88.5 | 5-shot | https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard |
| 14 | Qwen2.5-72B-Instruct | 88.2 | 5-shot | https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard |
| 15 | Qwen3-Coder-Next | 87.73 | From official technical report arXiv:2603 (accuracy) | https://arxiv.org/html/2603.00729v1 |
| 16 | Grok-2 | 87.5 | 5-shot | https://x.ai/blog/grok-2 |
| 17 | GPT-4 Turbo | 86.5 | 5-shot | https://openai.com/index/gpt-4-research/ |
| 18 | Qwen2.5-32B-Instruct | 86.1 | 5-shot | https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard |
| 19 | Qwen2.5-72B | 86.1 | 5-shot | https://qwenlm.github.io/blog/qwen2.5-llm/ |
| 20 | Llama 3.1 70B Instruct | 86 | 5-shot | https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard |
Interpretation
Trust when:
- Use top rows for shortlisting, then validate provider coverage and feature fit before selecting a model.
- Cross-benchmark consistency is a stronger signal than one-off score dominance.
Don't trust when:
- Don't compare raw score gaps across different benchmark scales as equivalent.
- Missing rows in the heatmap can hide benchmark blind spots.
- No trend data means this is a snapshot; expect drift in fresh refresh cycles.
Cross-benchmark heatmap
Compare top models across selected benchmark families to spot strengths and weak spots. On narrow screens, scroll horizontally to inspect every benchmark.
| Model | MMLU | GPQA | HumanEval | HellaSwag |
|---|---|---|---|---|
| Gemini 3.1 Pro Preview | 98.0 | 94.3 | 94.0 | — |
| GPT-5.5 | 92.4 | 93.6 | 94.2 | — |
| Claude Opus 4.6 | 91.1 | 91.3 | 95.0 | — |
| DeepSeek V4 Pro | 90.1 | 90.1 | 76.8 | — |
| Xiaomi MiMo-V2.5-Pro | 89.4 | 66.7 | — | — |
| Claude Sonnet 4.6 | 89.3 | 89.9 | 98.0 | — |