LiveBench
Contamination-resistant general capability, refreshed monthly by the LiveBench project.
23 models scored
Leaderboard →Benchmarks
Every leaderboard we score from.
The five public benchmarks behind the League Score — and every tracked model's result on each, with sources and dates.
SWE-bench Verified
Real GitHub issue resolution on the human-verified 500-task subset of SWE-bench.
13 models scored
Leaderboard →GPQA Diamond
Graduate-level science questions written to resist web lookup; Diamond is the hardest 198-question subset. Scores via the BenchLM GPQA leaderboard, whose ledger notes most rows are Diamond while some are GPQA main-set — treat close gaps as directional.
19 models scored
Leaderboard →MMLU-Pro
A harder, 10-option successor to MMLU across 14 subjects, ~12,000 questions.
8 models scored
Leaderboard →LMArena
Crowd preference Elo from blind head-to-head battles (formerly LMSYS Chatbot Arena).
10 models scored
Leaderboard →