AI benchmarks: the full set
Each benchmark has a leaderboard, a methodology note and links to the individual model-benchmark pages. Scores are modelled estimates on a 0–100 scale, except Arena Elo.
MMLU-Pro
Top score 91.6 by GPT-5.2 Pro. Average 73.5 across 600 models.
ReasoningGPQA Diamond
Top score 87.0 by GPT-5.2 Reasoning. Average 65.6 across 600 models.
CodingSWE-bench Verified
Top score 81.7 by GPT-5.2 Pro. Average 57.5 across 600 models.
CodingHumanEval
Top score 96.8 by GPT-5.2 Pro. Average 74.0 across 600 models.
MathematicsMATH-500
Top score 96.7 by GPT-5.2 Max. Average 72.5 across 600 models.
MathematicsAIME 2025
Top score 91.4 by Claude Opus 4.5 Pro. Average 63.6 across 600 models.
CodingLiveCodeBench
Top score 85.1 by Claude Sonnet 4.5 Reasoning. Average 61.0 across 600 models.
Instruction followingIFEval
Top score 96.1 by Claude Opus 4.5 Pro. Average 78.6 across 600 models.
SafetyTruthfulQA
Top score 91.1 by GPT-5.1 Max. Average 73.6 across 600 models.
CommonsenseHellaSwag
Top score 96.1 by GPT-5.2 Max. Average 81.4 across 600 models.
MultimodalMMMU
Top score 86.9 by GPT-5.2 Max. Average 67.1 across 600 models.
PreferenceArena Elo
Top score 1501.9 by Claude Opus 4.5 Max. Average 1349.9 across 600 models.