Every AI leaderboard

Leaderboards28Models ranked992

28 leaderboards · 992 models ranked

AI benchmarks: what is being measured

BenchmarkScores
MMLUAnswer four-option multiple-choice questions across 57 subjects, from elementary math to law.Knowledge · 2020UC Berkeley·
MMLU-ProAnswer ten-option multiple-choice questions that need more reasoning than MMLU.Knowledge · 2024TIGER-Lab (University of Waterloo)100
GPQA (Diamond)Answer graduate-level biology, physics and chemistry questions that non-experts cannot look up.Knowledge · 2023New York University100
Humanity's Last ExamAnswer expert-written closed-ended questions at the frontier of human knowledge, some with images.Knowledge · 2025Center for AI Safety and Scale AI44
SimpleQAAnswer short fact-seeking questions with a single correct answer, or decline.Knowledge · 2024OpenAI·
TruthfulQAAnswer questions where a common misconception tempts a false answer.Knowledge · 2021University of Oxford and OpenAI·
ARC-AGI-1Infer the rule behind a few input-output grid pairs and apply it to a new grid.Reasoning · 2019ARC Prize Foundation (created by François Chollet)·
ARC-AGI-2Solve harder abstract grid puzzles that are easy for people and hard for AI.Reasoning · 2025ARC Prize Foundation94
ARC-AGI-3Play unfamiliar interactive games with no instructions: explore, find the goal and win as efficiently as a person.Reasoning · 2026ARC Prize Foundation·
BIG-Bench HardSolve 23 BIG-Bench tasks where earlier models fell short of average human raters.Reasoning · 2022Google·
BIG-benchA broad collaborative suite of tasks believed to be beyond language models at the time.Reasoning · 2022Google and 450+ collaborators·
HellaSwagPick the most plausible continuation of an everyday scenario.Reasoning · 2019University of Washington and Allen Institute for AI·
AI2 Reasoning Challenge (ARC)Answer grade-school science multiple-choice questions; the Challenge set defeats simple retrieval.Reasoning · 2018Allen Institute for AI·
WinoGrandeResolve which of two options a pronoun-like blank refers to, using commonsense.Reasoning · 2019Allen Institute for AI and University of Washington·
MuSRSolve long natural-language murder mysteries, object placements and team allocations step by step.Reasoning · 2023University of Texas at Austin·
LiveBenchAnswer fresh questions in math, coding, reasoning, data analysis, language and instruction following.Reasoning · 2024LiveBench (Abacus.AI, NYU, NVIDIA and others)64
GSM8KSolve grade-school math word problems that take two to eight steps.Math · 2021OpenAI·
MATH (and MATH-500)Solve competition mathematics problems and give the final answer.Math · 2021UC Berkeley·
AIMESolve the American Invitational Mathematics Examination, a qualifier for the US Math Olympiad.MathMathematical Association of America·
FrontierMathSolve original, unpublished research-level mathematics problems with automatically checkable answers.Math · 2024Epoch AI·
MGSMSolve grade-school math word problems in ten typologically diverse languages.Math · 2022Google·
HumanEvalWrite a Python function from its signature and docstring.Coding · 2021OpenAI·
MBPPWrite short Python programs from a one-sentence description.Coding · 2021Google·
SWE-benchResolve a real GitHub issue by editing the repository so the hidden tests pass.Coding · 2023Princeton University·
SWE-bench VerifiedResolve real GitHub issues; the subset people confirmed is well specified and fairly tested.Coding · 2024SWE-bench team and OpenAI39
SWE-Bench ProSolve long-horizon, enterprise-style software engineering tasks in real repositories.Coding · 2025Scale AI18
LiveCodeBenchSolve competitive programming problems published after a model was trained; also self-repair and test-output prediction.Coding · 2024UC Berkeley, MIT and Cornell·
Aider PolyglotEdit code to solve hard programming exercises in six languages, through the Aider coding assistant.Coding · 2024Aider50
Terminal-BenchComplete hard, realistic tasks in a computer terminal: build, debug, configure, analyse.Coding · 2025Stanford University and Laude Institute·
SciCodeWrite code for research problems from 16 natural-science subfields.Coding · 2024University of Illinois and collaborators·
BigCodeBenchWrite Python that calls many library functions to complete practical tasks.Coding · 2024BigCode·
MLE-benchAct as a machine-learning engineer on Kaggle competitions: prepare data, train, submit.Coding · 2024OpenAI·
SWE-LancerComplete real freelance software jobs from Upwork, both implementation and choosing between proposals.Coding · 2025OpenAI·
τ-bench (tau-bench)Serve a simulated customer through tools while following a domain policy.Agents and tools · 2024Sierra·
Berkeley Function Calling Leaderboard (BFCL)Call functions and tools correctly: single, parallel, multi-turn, and agentic web search and memory.Agents and tools · 2024UC Berkeley (Gorilla)81
MCP AtlasComplete realistic tasks by calling tools on real MCP servers, often across several servers.Agents and toolsScale AI37
GAIAAnswer real-world questions that need browsing, tools, files and several steps.Agents and tools · 2023Meta, Hugging Face and AutoGPT·
WebArenaComplete tasks on realistic self-hosted websites: shopping, forums, code hosting, maps, content management.Agents and tools · 2023Carnegie Mellon University·
OSWorldOperate a real computer (Ubuntu, Windows, macOS) through screenshots and input to finish open-ended tasks.Agents and tools · 2024University of Hong Kong (XLang)·
BrowseCompFind hard-to-locate, entangled facts by persistently browsing the web.Agents and tools · 2025OpenAI·
GDPvalProduce real work products (documents, slides, spreadsheets) for tasks from 44 occupations.Agents and tools · 2025OpenAI·
Vending-BenchRun a simulated vending machine business for a long time: inventory, orders, prices, fees.Agents and tools · 2025Andon Labs·
IFEvalFollow verifiable formatting instructions such as "write more than 400 words".Chat and instructions · 2023Google·
MultiChallengeHold realistic multi-turn conversations: retain instructions, remember context, stay self-coherent.Chat and instructions · 2025Scale AI35
RULERRetrieve, trace and aggregate information across long synthetic contexts, beyond one needle.Long context · 2024NVIDIA·
LongBenchUnderstand long real documents in English and Chinese: QA, summarisation, code, few-shot.Long context · 2023Tsinghua University and Zhipu AI·
∞Bench (InfiniteBench)Answer questions over contexts longer than 100,000 tokens: books, code, math, dialogue.Long context · 2024Tsinghua University and collaborators·
MRCR (multi-round coreference)Find the right one of several near-identical requests hidden in a long conversation and reproduce its answer.Long context · 2024Google DeepMind; open version by OpenAI·
MMMUAnswer college-level questions that need an image: charts, diagrams, maps, music sheets, chemical structures.Multimodal · 2023MMMU team·
MathVistaSolve math problems set in images: plots, tables, geometry, puzzles.Multimodal · 2023UCLA, University of Washington and Microsoft·
ChartQAAnswer questions about charts that need visual and logical reasoning.Multimodal · 2022York University and Nanyang Technological University·
DocVQAAnswer questions about scanned document images.Multimodal · 2020CVC Barcelona and IIIT Hyderabad·
Video-MMEAnswer questions about videos from 11 seconds to an hour long, with or without subtitles and audio.Multimodal · 2024Video-MME team·
MMBenchAnswer multiple-choice questions testing 20 vision-language abilities.Multimodal · 2023Shanghai AI Laboratory and collaborators·
AI2DAnswer questions about grade-school science diagrams.Multimodal · 2016Allen Institute for AI·
GenEvalGenerate images that get objects, counts, colours and positions right.Image and video generation · 2023University of Washington·
VBenchGenerate videos that score well on 16 separate quality and consistency dimensions.Image and video generation · 2023Nanyang Technological University and Shanghai AI Laboratory·
Open ASR LeaderboardTranscribe English short- and long-form speech and multilingual speech.SpeechHugging Face and collaborators63
LibriSpeechTranscribe read English audiobook speech.Speech · 2015Johns Hopkins University·
Common VoiceTranscribe crowd-sourced read speech in many languages.Speech · 2019Mozilla·
FLEURSRecognise, identify and translate speech in 102 languages.Speech · 2022Google·
Artificial Analysis Speech ArenaRead text aloud in a voice people prefer.SpeechArtificial Analysis100
MTEB (and MMTEB)Embed text for retrieval, clustering, classification, reranking and similarity.Embeddings · 2022MTEB community (Hugging Face and collaborators)96
HarmBenchRefuse harmful behaviours under automated red-teaming attacks.Safety and honesty · 2024Center for AI Safety·
AgentHarmRefuse explicitly malicious multi-step agent tasks that use tools.Safety and honesty · 2024UK AI Security Institute and Gray Swan AI·
StrongREJECTMeasure how much a jailbreak really helps: refusals and the usefulness of any harmful answer.Safety and honesty · 2024UC Berkeley·
XSTestAnswer safe prompts that only look risky, and refuse the unsafe contrasts.Safety and honesty · 2023Bocconi University and University of Oxford·
BBQAnswer questions where a social stereotype could bias the answer.Safety and honesty · 2021New York University·
WMDPMeasure hazardous knowledge in biosecurity, cybersecurity and chemical security (lower can be safer).Safety and honesty · 2024Center for AI Safety and Scale AI·
MASKTell whether a model states what it believes when pressured to lie, separately from whether it is accurate.Safety and honesty · 2025Center for AI Safety and Scale AI·
LMArena Text Arena (Chatbot Arena)Chat: people ask anything and vote for the better of two anonymous answers.Arenas · 2023LMArena (from LMSYS, UC Berkeley)84
LMArena WebDev ArenaBuild a web app from a prompt; people vote for the better of two.ArenasLMArena84
LMArena Vision ArenaChat about images; people vote for the better answer.ArenasLMArena85
LMArena Document ArenaAnswer questions about uploaded documents and PDFs; people vote.ArenasLMArena40
LMArena Search ArenaAnswer with web search grounding; people vote for the better answer.ArenasLMArena40
LMArena Text-to-ImageGenerate an image from a prompt; people vote for the better of two.ArenasLMArena71
LMArena Image EditEdit an image from an instruction; people vote for the better edit.ArenasLMArena51
LMArena Text-to-VideoGenerate a video from a prompt; people vote for the better of two.ArenasLMArena46
LMArena Image-to-VideoAnimate an image into a video; people vote for the better of two.ArenasLMArena45
Artificial Analysis Intelligence IndexOne number from ten independent evaluations that Artificial Analysis runs itself.Composite indexesArtificial Analysis54
Epoch Capabilities IndexOne capability score stitched together from dozens of benchmarks.Composite indexesEpoch AI100
Open LLM LeaderboardRank open-weights models on a fixed set of academic benchmarks run the same way.Composite indexes · 2023Hugging Face·
Artificial Analysis Output SpeedMeasure how fast a model streams tokens on its first-party API.Speed, price and usageArtificial Analysis46
Artificial Analysis PriceCompare what a model costs on its first-party API.Speed, price and usageArtificial Analysis19
OpenRouter usage rankingsShow which models developers actually send tokens to.Speed, price and usageOpenRouter19

Scores counts the models with a captured score on the boards that publish this benchmark. A dot means we capture no board for it yet.

Sources: each benchmark's paper (arXiv) and its official site or leaderboard. Scores are the boards captured here every day, exactly as published.