Rankings

Evaluate AI systems on open-ended investment research.

Compare vision models on detection, counting, identification, OCR, and visual reasoning.

Evaluation of emotional intelligence in real multi-turn human-AI conversations.

A benchmark of coding models finding and fixing hidden bugs in production codebases.

Evaluation of agents performing stateful tool-use and multi-app workflows.

Question answering over document images, combining reading with document-layout understanding.

Grounded question answering about publicly traded companies using financial documents.

A collaborative collection of tasks for evaluating legal reasoning in language models.

Evaluate models on practical scientific reasoning and life-science research work.

An informal comparison of language models asked to draw a pelican riding a bicycle using SVG code.

Track AI models making forecasts through simulated prediction-market trades.

Evaluate research-level mathematical reasoning on expert-created problems.

Evaluate coding agents on resolving real software repository issues.

A catalog of evaluations for agents, professional work, reasoning, and writing.

Test multi-step numerical reasoning over financial reports.

Test numerical reasoning across conversations about financial reports.

Evaluate financial language models across document analysis, forecasting, and decisions.

Compare language models across financial questions and banking intents.

Evaluate agents researching public-company filings and financial analysis.

Test agents on extended finance assignments with realistic files and deliverables.

Test whether agents can reason through professional spreadsheets.

Evaluate strategic reasoning through competitive economic games.

Evaluate model responses to realistic health conversations with physician-written rubrics.

Test medical knowledge and reasoning using professional medical-board exam questions.

Compare models on medical entrance-exam questions across subjects and healthcare topics.

Answer biomedical research questions using evidence from corresponding PubMed abstracts.

Evaluate agents completing clinical tasks in a virtual electronic-health-record environment.

Evaluate tool-using medical agents through simulated patient encounters and multimodal clinical evidence.

Compare language models across clinical support, documentation, patient communication, research, and administration.

Test agents on extended healthcare assignments with brief handoffs and complex supporting materials.

Evaluate how precisely AI retrieves evidence from legal documents.

Compare language models on legal judgments, legislation, and contracts.

Evaluate AI on professional deliverables across major industries.

Compare customer-service agents handling retail, travel, and telecom requests.

Test agents on cybersecurity challenges in controlled environments.

Test whether tool-using agents complete tasks despite prompt-injection attacks.

Compare agents building machine-learning solutions for data-science competitions.

Compare computer-use agents on extended professional tasks.

Evaluate assistants answering questions that require research, tools, and multiple kinds of evidence.

Compare multimodal models on college-level questions across disciplines.

Test advanced scientific reasoning with expert-written questions.

Evaluate broad academic and professional knowledge across 57 subjects.

Test multi-step arithmetic reasoning through grade-school word problems.

Evaluate whether generated Python functions pass their tests.

Compare AI writing through judgments from professional writers.

Evaluate agents that turn scientific data and instructions into working analysis programs.

Assess literature, data, and protocol reasoning used in biology research.

Test code for scientific calculations, numerical methods, and simulations.

Compare agents completing difficult tasks in a terminal environment.

Evaluate freelance software fixes and choices between implementation proposals.

Evaluate whether generated CAD designs can be edited, assembled, and checked mechanically.

Measure security agents against known web-application vulnerabilities in isolated test environments.

Test whether an agent can reproduce a documented software vulnerability.

Assess interpretation of malware-analysis evidence and threat-intelligence reports.

Evaluate agents that repair security defects in native code.

Judge explanations, feedback, and hints tailored to a learner.

Assess how tutors respond to student mistakes and confusion in mathematics.

Test agents that search for and select products matching a shopper’s requirements.

Evaluate comparison shopping across stores, including compatible products and checkout.

Test visually guided web tasks, including product selection and shopping actions.

Evaluate writing against requirements specific to each assignment.

Check whether models can sustain useful writing at a requested length.

Compare short creative writing and inspect the generated samples.

Compare embedding models for search, classification, clustering, and related tasks.

Evaluate how well embedding models find relevant documents on unseen retrieval tasks.

Compare speech-to-text models on transcription accuracy and processing speed.

Evaluate Python programs that combine libraries and follow complex instructions.

Find model and agent evaluations published by research teams and the Kaggle community.

Compare AI models playing strategic games against one another.

Compare coding models on dated contest problems and code-related reasoning tasks.

Evaluate whether models use tools correctly across calls, conversations, search, and memory.

Evaluate agents querying and transforming data in complex enterprise databases.

Recreate a webpage in HTML and CSS from a screenshot.

Compare models writing and editing code across six programming languages.

Test everyday reasoning, social understanding, and resistance to misleading wording.

Infer unfamiliar visual rules from examples and apply them to new grids.

Test whether assistants remember, update, and reason about information across conversations.

Evaluate generated videos for subject consistency, motion, visual quality, and prompt alignment.

Test challenging academic knowledge and reasoning with harder multiple-choice questions.

Evaluate difficult academic questions across mathematics, science, and the humanities.

Evaluate reading text in images, answering document questions, and extracting key information.

Extract complete, structured records and supporting evidence from enterprise documents.

Evaluate AI agents creating motion graphics through code that renders into video.

Compare factual accuracy in image questions, world knowledge, search, and document-grounded responses.

Compare how much code coding agents add while solving the same tasks.

Compare AI-designed product screens with corresponding real product designs.

Test how accurately and completely agents research entities and assemble sourced answers.

Compare search APIs on scholarly, news, financial, legal, and obscure-entity queries.

Evaluate summarization, extraction, correction, and dictation from transcripts.

Test whether agents recover interrupted wet-lab experiments and protect time-sensitive samples.

Compare agents keeping research, spreadsheets, an app, and recommendations consistent as requirements change.

Compare machine-learning methods for biomedical tables with many variables and limited samples.

Test whether coding agents preserve working behavior and code quality as requirements evolve.

Evaluate the path from an application specification to a tested cloud deployment.

Evaluate agents completing research workflows contributed by scientists.

Evaluate ingredient substitution, pairing, and constrained culinary choices with executable scoring.

Test planning and tool use by having coding agents play RuneScape.

Compare code-review agents on valid findings, missed issues, and review noise.

Test whether language models write correct GPU kernels that outperform reference implementations.

Compare predictive machine-learning methods on curated tabular datasets.

Compare language models coaching opposing Blood Bowl teams.

Test recognition and reading of ancient Egyptian hieratic handwriting.

Test whether coding agents can discover and repair bugs without being told where to look.

Compare playable Pac-Man games generated from a single short prompt.

Compare raw language models reviewing the same real pull requests.

Compare generated motion graphics, rendering success, cost and time.

Test whether coding agents can reconstruct dynamic scenes from video.

Evaluate retrieval over varied company documentation and knowledge.

Compare computer-use agents on task success, speed and cost.

Test whether agents can select the next experiment from real wet-lab records.

Evaluate agents diagnosing and operating real AWS infrastructure.

Track accepted AI-assisted solutions to formal mathematics problems in Lean.

Compare AI models and API providers on quality, price, speed, and latency.

Compare 300+ AI models on public benchmark results, price, and speed.

Compare AI models across hundreds of public benchmarks with linked sources.

Track frontier model results across benchmarks Epoch runs or collects.

Rank models from anonymous, crowdsourced pairwise votes.

Browse Scale AI’s expert-built benchmarks for agents, reasoning, and safety.

Compare recent frontier models on unsaturated public benchmarks, cost, and speed.

Compare language models on Japanese generation, application development, and safety.

Rank AI-generated websites, apps, slides, images, video, and audio from pairwise votes.

Compare open-weight language models on a fixed automatic benchmark suite.

Compare language and multimodal models on OpenCompass evaluation suites.

Evaluate language models across many scenarios and metrics with a shared framework.

Run every model on samples from the same set of public benchmarks and average the results.

Compare AI agents on finance, coding, legal, and tax work, weighted by economic share.

Compare language, multimodal, and other foundation models on BAAI’s evaluation platform.

Rank language models on one scale fused from many public benchmark leaderboards.

Compare commercial and open models on a continuously updated Chinese-language evaluation.

Browse a curated table of AI leaderboards by category.

Browse a large curated list of LLM evaluation benchmarks, tools, and leaderboards.

Test commonsense inference by choosing the most plausible ending to an everyday situation.

Resolve ambiguous pronouns that require commonsense knowledge.

Test physical commonsense by choosing the better way to accomplish a goal.

Answer grade-school science questions that resist simple retrieval.

Answer science questions that combine a given fact with broader common knowledge.

Answer questions that require everyday commonsense knowledge.

Answer naturally occurring yes/no questions about a passage.

Answer reading questions that need counting, sorting, or arithmetic over a paragraph.

Answer trivia questions using retrieved web or Wikipedia evidence.

Answer real search questions from a full Wikipedia page.

Evaluate general language understanding across eight harder tasks.

Probe language models on more than 200 crowd-contributed tasks.

Solve competition mathematics problems across seven subjects.

Solve grade-school math word problems in ten languages.

Write short Python functions from natural-language descriptions.

Test models on questions from human admission and qualification exams.

Evaluate Chinese knowledge and reasoning across 52 disciplines.

Reason through long narratives such as murder mysteries and team assignments.

Check whether models follow verifiable formatting and content instructions.

Judge multi-turn conversation quality with a strong model as grader.

Estimate instruction-following quality with an automatic judge.

Measure whether models answer short factual questions correctly or decline.

Detect when safety training makes models refuse safe requests.

Test model refusals against risk categories drawn from regulations and company policies.

Evaluate Chinese-language alignment and answer quality.

Test code generation across many programming languages.

Answer multimodal science questions and explain the reasoning.

Evaluate retrieval-augmented answers to realistic questions across five domains.

Test how well systems detect factual errors in model responses.

Test knowledge and application of Chinese law.

Infer people’s goals and beliefs from video and text.

Solve Olympiad-level problems across seven science disciplines.

Solve college-level problems in physics, chemistry, and mathematics.

Answer questions that combine tables and text from financial reports.

Answer questions over financial report pages with tables and text.

Diagnose how multimodal models solve visual math problems step by step.

Explain what is wrong with images that defy common sense.

Test multi-step reasoning over images and text across domains.

Answer real-world programming questions in the style of Stack Overflow.

Measure honesty: self-knowledge, non-deception, and consistency.

Answer questions using knowledge bases, text, tables, and infoboxes together.

Evaluate personalized image generation with human-aligned automatic judging.

Approximate crowd preference rankings cheaply with mixed ground-truth questions.

Answer medical exam questions in six languages.

Test comprehension of ancient Chinese.

Test models on regularly refreshed tasks graded against objective answers.

Compare evaluation methods for language models with large-scale human and model judging.

Compare models’ mathematical problem solving across many datasets.

Compare large language models on Tsinghua’s periodic capability evaluation.

Evaluate Swedish language understanding across many tasks.

Solve the American Invitational Mathematics Examination problems for a given year.

Evaluate emotional and social intelligence in multi-turn conversations.

Test coding agents on long-horizon changes in professional repositories.

Evaluate coding agents on codebase questions, test writing, and refactoring.

Test real-world tool use through Model Context Protocol servers.

Test whether agents ask for help when information is missing or ambiguous.

Evaluate coding agents on early-stage drug discovery work.

Measure whether AI agents can complete paid freelance projects end to end.

Test intuitive visual reasoning about what images and videos imply.

Evaluate spoken dialogue systems over long, natural multi-turn conversations.

Test whether models handle realistic multi-turn conversations correctly.

Grade professional finance and legal reasoning with expert rubrics.

Measure model safeguards against national-security and public-safety misuse.

Evaluate the quality of crisis support models give people in distress.

Measure coding agents on original, long-horizon engineering tasks.

Evaluate coding agents on fresh tasks collected after model release dates.

Judge whether coding agents produce changes a maintainer would merge.

Evaluate coding agents on ambiguous, multi-file tasks from real Cursor sessions.

Compare how well models write working code for the Convex backend platform.

Measure how well models find security vulnerabilities in application code.

Test advanced mathematical reasoning on unpublished, expert-written problems.

Compare embedding models across text, images, video, audio, and documents.

Compare short stories written to the same constrained creative brief.

Evaluate planning, revising, and writing a novella-length story.

Evaluate typed yes/no, choice, and score decisions across applied domains.

Compare decision models on English tasks with a fixed sample.

Test whether a decision model’s probabilities stay coherent across related questions.

Compare models on composite capability, cost, and decision-model benchmarks.

Test whether agents complete cross-application business workflows through APIs.

Test whether models handle real professional PDFs and prompts.

Test whether models read specialized professional charts correctly.

Evaluate legal agents on long-horizon client-matter work.

Evaluate research agents on finding complete answer lists across sources.

Test coding agents on ultra-long-horizon engineering and research tasks.

Evaluate coding agents on building complete web applications.

Test whether agents can post-train small language models within a compute budget.

Test long-context reasoning by following edges in large graphs.

Evaluate agents on long-horizon professional workflows with verifiable outcomes.

Test understanding of very long videos.

Test whether coding agents can find and patch undisclosed software weaknesses.

Answer diverse, harder questions about real-world charts.

Evaluate agents on spreadsheet-centric finance and accounting workflows.

Test how well the model inside an agent resists adversarial attacks.

Evaluate how AI models behave in conversations with children and teens.

Track jailbreak attacks and defenses with a shared benchmark.

Score how useful a jailbroken model’s harmful answers actually are.

Measure whether agents carry out harmful multi-step tasks with tools.

Evaluate agents on physics-checked electrical engineering tasks.

Measure how well voice agents detect when a speaker has finished talking.

Test agents on interactive reasoning games they have never seen.

Test reasoning across long documents of 10,000 to 100,000 tokens.

Measure factual knowledge and hallucination across domains.

Evaluate agents on multi-week business projects with real deliverables.

Run OpenAI’s GDPval occupational tasks with agents in Artificial Analysis’s harness.

Test whether models follow new, unseen output constraints.

Test models on unpublished, research-level physics problems.

Measure short-form factual recall with a corrected question set.

Evaluate web research agents on realistic research tasks.

Evaluate agentic reasoning across text and visual games.

Test agents on unusual machine-learning tasks with limited data and feedback.

Test understanding of long stories at increasing context lengths.

Test whether speech recognition keeps exact values such as identifiers, paths, and commands.

Pit two AI agents against each other, each attacking and defending changing services at once.

Describe how models write and communicate by default, from warmth to confidence to emoji use.

Test how accurately and consistently AI agents solve data-analysis problems.

Test whether AI agents can analyze obfuscated malware and extract useful evidence.

Test reasoning through spoken questions and spoken answers.

Test whether AI agents can improve on published computer-science research code.

Compare accuracy, cost, and speed when the same open-weight model runs on different providers.

Test whether multimodal agents can control simulated robots across varied tasks and bodies.

Compare whether AI coding agents produce code that both works and passes security tests.

Compare text-to-speech models through listener preferences using the same cloned voices.

Compare text-to-speech models using each provider’s own voices and listener preferences.

Compare game-playing policies on deterministic Drop7 rounds with matched hidden randomness.

Evaluate image-based 3D reconstruction against laser-scanned indoor and outdoor scenes.

Evaluate tool use across routine calls, imperfect instructions, and multi-turn agent interactions.

Evaluate recognition and detection of everyday human activities in videos.

Test whether analytics agents answer business questions while catching ambiguity, statistical traps, and data defects.

Evaluate real-time decisions by models navigating a Pac-Man game.

Evaluate whether adversarial inputs defeat language-model safeguards.

Evaluate complex, multi-turn, and system-level instruction following with expert-written rubrics.

Evaluate scientific code using an independently corrected edition of SciCode tasks and grading tests.

Evaluate natural-language inference on examples written to expose model weaknesses.

Evaluate advanced coding and algorithmic reasoning with problems from programming competitions.

Evaluate language-model agents reasoning and acting across interactive environments.

Evaluate coding agents on real repository changes with held-out tests.

Compare AI gateway caching, routing, latency, and cost under repeated agent workloads.

Compare defects found by AI code reviewers and runtime verification.

Evaluate diagram structure understanding and question answering about scientific illustrations.

Evaluate agents on professional tax assignments with complex records and incomplete instructions.

Evaluate smartphone GUI agents using screen observations, action decisions and intermediate reasoning.

Test mathematical reasoning on the AMC contest selection documented in the Gemini 1.5 evaluation.

Test advanced mathematical reasoning on original problems at Olympiad difficulty or above.

Evaluate coding agents on Android app development, migrations and architectural changes.

Compare web-search APIs on focused queries for agent workflows in sales, insurance and finance.

Compare model knowledge and reasoning on a fixed sample of expert-level multiple-choice questions.

Evaluate whether agents can follow instructions and generalize across Android apps and tasks.

Test agents discovering, reproducing and patching vulnerabilities in real software projects.

Compare document-editing tools used by an AI agent on legal work from Harvey LAB.

Compare automatic PCB routing tools on real open-hardware board designs.

Evaluate coding agents on real changes to Snowflake’s large Java, C++ and Go codebases.

Test whether code-review agents detect bugs that caused real PostHog incidents.

Evaluate chemistry assistance and risk awareness when questions have both beneficial and harmful uses.

Measure whether AI chat systems avoid hazardous responses to direct requests.

Test whether AI safeguards withstand adversarially rewritten requests.

Compare agent retrieval accuracy, token use and latency across SQLite and Markdown context stores.

Can an agent finish the work that customers already delegate to an AI startup?

Test whether an image model can reproduce a specified typeface, match a requested color and make a targeted edit.

How close is a model’s location estimate when the photo has never been online?

Track whether a deployed model still completes the same work as efficiently as it did at launch.

Do shopping assistants recommend brands associated with a requested style, or drift into a neighbouring one?

Can a language model solve a scrambled cube given only its text representation?

Compare voice agents over a conversation long enough for per-turn behavior and costs to change.

Compare model capabilities, serving performance and readiness for agent workflows in one evaluation harness.

Evaluate forecasting models across varied time scales, domains and prediction horizons.

Compare forecasting models on realistic tasks, including forecasts that use additional explanatory variables.

Test zero-shot forecasting on tasks configured around real operational needs.

Compare how accurately speech models recognize emotion across 29 languages.

Compare three decisions a voice assistant makes: when to speak, which action to take, and how to deliver a reply.

Test whether an agent can complete tasks in working Android apps by checking what actually changed on the device.

Test whether an agent can turn a known software flaw into a working exploit in an isolated target.

Test whether a security agent can demonstrate a known vulnerability against a remotely accessible application in a controlled environment.

Test whether a coding agent can keep a web application working as new feature requests arrive.

See what happens when people turn language models’ instructions and the wood they have on hand into chairs.

Test Chinese language models on civil-service exam questions and see which reasoning skills account for their mistakes.

Test whether an agent can inspect a working application and build a replacement that behaves and looks like it.

Compare how often decision models miss phishing emails or flag legitimate mail, and how those errors change with the share of phishing in an inbox.

Compare text classifiers and language models on sentiment, news topics and banking requests as the number of possible labels rises from two to 77.

Evaluate a model’s ability to produce professional work from a brief and source documents in banking, consulting, law and primary care.

Compare Claude Sonnet 4.6 with the small Strands Decider model for sorting tickets in an internal Amazon system into 20 categories, by cost and response time.

Compare how well AI agents complete specialist work, from designing chips to analyzing financial records, and how much each attempt costs.

See whether a shopping agent can fill a grocery cart that fits a shopper’s budget and tastes while limiting waste.

Measure which facts survive when language models repeatedly turn an answer into a question and try to reconstruct the answer.

Test whether a coding agent can sustain progress on a long software project and verify its work before stopping.

Evaluate whether an agent can close the books for a simulated company using accounting software, spreadsheets and supporting documents.

Evaluate whether a voice agent completes an enterprise task accurately while sustaining a natural spoken conversation.

Test agents on extended banking, consulting and legal work spread across files, communications and applications.

Compare native voice models across reasoning, customer-service tasks, human preference and successful tool use.

Compare human preferences and successful task completion in live calls with anonymous voice models.

Separate an agent’s task performance from the reliability of the software harness running it.

Test mathematical reasoning on a small set of competition problems selected for difficulty.

Test whether coding agents reproduce clinical variables from electronic health records.

Test agents on building integrated software systems and diagnosing production-style failures.

Can an agent leave cloud infrastructure in the state an operator requested?

How much Buddhist history do language models know across regions and historical periods?

How many real bugs can a QA agent find in a running web app, and how often are its reports accurate?

Can a web research agent find a large set of qualifying records and support each claim with evidence?

How well does a forecasting agent revise an event probability when new reporting becomes available?

How often does a detector find personal information in text and assign it the correct type?

Can a model reconstruct parts, assemblies and circuit boards from engineering drawings and rendered images?

Can a robot manipulation policy remember earlier observations, handle unfamiliar objects and finish a sequence of precise movements?

Compare AI video generators for character consistency, action scenes, and creative work.

Compare financial research grounded in filings, calculations, and evidence.

Compare AI support for market research, startup diligence, and investment memos.

Compare AI on medical questions, evidence review, and clinical documents.

Compare reading accuracy and structured extraction from scans, tables, and difficult layouts.

Compare practical coding work through working changes, independent tests, and clear review.

Compare how faithfully AI rebuilds an editable webpage from its screenshot.

Compare AI conversation partners for people practicing a new language.

Compare AI that labels tickets, routes requests, and tags documents, from large language models to small, fast decision models.