Benchmark leaderboards
Each standard test has its own leaderboard, with an explanation of what it measures, how it is scored, and the source of every result. Only tests with at least five published scores get a page.
Reasoning
- GPQA DiamondGraduate-level, Google-proof science questions in biology, physics, and chemistry.GPT-6 Astra 96.1 · 19 models
- Humanity's Last ExamExpert-written closed-ended questions at the frontier of human knowledge.Claude Opus 5.5 64.4 · 23 models
- AIMEAmerican Invitational Mathematics Examination competition problems.Reflection Beam 1.0 97.7 · 5 models
Coding
- LiveCodeBenchFresh competitive-programming problems collected after each model's training cut-off.DeepSeek V4 Pro 93.5 · 7 models
- SWE-bench VerifiedReal GitHub issues from popular Python repositories, fixed end to end.Reflection Beam 1.0 81.6 · 5 models
- SWE-bench ProHarder, longer software tasks across many languages and repositories.Claude Opus 5.5 89.9 · 9 models
- Terminal-BenchHands-on tasks completed inside a real Linux terminal.DeepSeek V4.1 Flash 90.6 · 13 models
- Terminal-Bench 4.0The newest, much harder Terminal-Bench task set.Claude Sonnet 5.5 70.6 · 16 models
Agentic
Knowledge
Human preference
Scores are published results from model cards, lab reports, and public leaderboards, each linked to its source. Settings such as thinking mode or test version differ between sources and are shown next to each score. OpenLLM Buddy catalog models are uncensored fine-tunes; their scores are those of the official base model. Data collected 2026-10-30. How the index works