AI benchmarks
What each benchmark measures, with links to the original source. We don't republish scores, because they change quickly and depend on the setup.
See leaderboards| Benchmark | Measures | Source |
|---|---|---|
| MMLU | Knowledge across 57 subjects | Open ↗ |
| MMLU-Pro | Harder, reasoning-focused MMLU | Open ↗ |
| GPQA | Graduate-level, Google-proof science Q&A | Open ↗ |
| Humanity's Last Exam | Frontier expert-level questions | Open ↗ |
| SWE-bench | Fixing real GitHub issues | Open ↗ |
| HumanEval | Python code generation | Open ↗ |
| LiveCodeBench | Contamination-free coding | Open ↗ |
| MATH | Competition mathematics | Open ↗ |
| AIME | American Invitational Mathematics Exam problems | Open ↗ |
| ARC-AGI | Abstract reasoning puzzles | Open ↗ |
| LMArena | Crowdsourced head-to-head chatbot ratings | Open ↗ |
| MMMU | Multimodal college-level reasoning | Open ↗ |
| TruthfulQA | Avoiding common falsehoods | Open ↗ |
| GAIA | General AI assistants with tools | Open ↗ |
| τ-bench | Tool-agent-user interaction | Open ↗ |
| OSWorld | Computer-use agents in real OSes | Open ↗ |
| BIG-bench | Diverse 200+ tasks | Open ↗ |
| HELM | Holistic evaluation of LMs | Open ↗ |
| Open LLM Leaderboard (archived) | Open-model evaluations | Open ↗ |
| IndicGenBench | Generation in Indian languages | Open ↗ |