// benchmark hub
AI benchmarks, explained in plain English
What each major AI benchmark actually tests, where it falls short, and where to find current, independently run scores. We link to the original authors and to Artificial Analysis for numbers rather than copying them.
- MMLU-Pro: harder, 10 option successor to MMLU across 14 subjects.
- GPQA Diamond: graduate level science questions written to be 'Google-proof'.
- Humanity's Last Exam (HLE): expert written questions at the frontier of human knowledge.
- SWE-bench (Verified): can a model fix real GitHub issues in real Python projects?
- LiveCodeBench: fresh competitive programming problems to limit training data leaks.
- AIME: US high school olympiad qualifier maths problems with integer answers.
- Terminal-Bench: can an AI agent complete real tasks in a Linux terminal?
- τ²-Bench (tau2): customer service agent tasks with a simulated user and real policies.
- MMMU / MMMU-Pro: college level questions that need both images and text.
- ARC-AGI: visual grid puzzles that test learning new rules from a few examples.
- Artificial Analysis Intelligence Index: a composite score that combines many evaluations into one number.
Reading benchmarks sensibly
- A score is only comparable when the same version of the test was run the same way (same prompts, same number of attempts, same tools allowed).
- Public test questions leak into training data over time, so older benchmarks get easier for reasons that aren't real skill.
- No single benchmark predicts how a model will do on your task. Try your own examples.