A score only means something if you know what was tested. Here is what each benchmark asks a model to do, who runs it, and who leads it today.
Artificial Analysis
One composite score of general capability, built from ten evaluations spanning agents, coding, general knowledge and scientific reasoning.
Current leaderClaude Opus 5.5
57.629 of 29 models scored
Epoch AI
A statistical composite that folds a model's results on dozens of benchmarks into one general-capability number.
Current leaderClaude Opus 5.5
167.316 of 29 models scored
LiveBench
A broad test whose questions are regularly replaced, so answers cannot simply have been memorised from training data.
Current leaderClaude Fable 5.1
83.4%23 of 29 models scored
Terminal-Bench team
Realistic command-line jobs in a Linux container — build, debug, analyse, administer — each checked by tests.
Current leaderClaude Opus 5.5
64.8%13 of 29 models scored
Scale AI
Long, multi-file fixes and features in real repositories, on the subset that still separates frontier coding agents.
Current leaderClaude Fable 5.1
92.2%5 of 29 models scored
Arena (formerly LMArena)
Blind head-to-head votes on web apps that two anonymous models build from the same prompt.
Current leaderClaude Opus 5.5
1,81427 of 29 models scored
Artificial Analysis
Real knowledge-work assignments from 44 occupations — reports, spreadsheets, slide decks — judged head to head.
Current leaderClaude Opus 5.5
1,86618 of 29 models scored
Andon Labs
Run a simulated vending-machine business for a year; the score is the money left in the bank.
Current leaderGPT-6 Astra
$15,5159 of 29 models scored
ARC Prize Foundation
Visual grid puzzles that require inferring a brand-new rule from a handful of examples.
Current leaderGPT-6 Astra
95%11 of 29 models scored
ARC Prize Foundation
Unfamiliar turn-based games with no instructions: explore, work out the goal, then win efficiently.
Current leaderGPT-6 Astra
62.7%5 of 29 models scored
Epoch AI
Original, unpublished mathematics problems from advanced undergraduate to early research level.
Current leaderGPT-6 Astra
93.7%14 of 29 models scored
SRI Lab, ETH Zurich and INSAIT
Recent mathematics competitions and olympiads, chosen so the problems post-date each model's training.
Current leaderGPT-6.1 Sol
90.1%10 of 29 models scored
SimpleBench team
Everyday trick questions about space, time and people that humans find easy and language models have tended to fumble.
Current leaderClaude Opus 5.5
88.4%12 of 29 models scored
Scale AI and the Center for AI Safety
A thousand expert-written questions at the edge of academic knowledge, across more than a hundred subjects.
Current leaderGPT-6 Astra
60.6%9 of 29 models scored
Epoch AI
Graduate-level multiple-choice science questions that experts get right and skilled non-experts get wrong.
Current leaderGPT-6 Astra
95.8%14 of 29 models scored
Arena (formerly LMArena)
Blind head-to-head votes on answers to prompts that include an image.
Current leaderQwen3.8-Max
1,30115 of 29 models scored
Arena (formerly LMArena)
Blind head-to-head votes from millions of real conversations: which answer did people prefer?
Current leaderGemini 4 Argon
1,52524 of 29 models scored