Reference

Benchmarks

A benchmark is a fixed test that lets different AI systems be compared on the same footing. This is a catalogue of the ones that have mattered — what each measures, who built it, where to find it, and how the leading score has moved over time. A recurring pattern runs through almost all of them: a benchmark is released to be hard, models climb it within a year or two, and it saturates — at which point a harder one is built to replace it. That cycle is followed as a story in the benchmarks and their saturation thread.

111 benchmarks · 12 domains

Reasoning & problem-solving

General reasoning, hard exams, puzzles that resist memorisation.

  • ARC-AGILive

    Whether a system can infer an unfamiliar abstract rule from a handful of examples and apply it to a new case, rather than recognising a pattern it has seen before.

    ARC Prize Foundation (François Chollet, Mike Knoop)5 November 2019Leader: Gemini 3 Deep Think (v2) (84.6% (ARC-AGI-2, verified))

  • Humanity's Last ExamLive

    Whether a model can answer the hardest closed-ended questions expert academics could write in their own field, at a difficulty chosen specifically to be far from saturated.

    Center for AI Safety & Scale AI23 January 2025Leader: Gemini 3 Deep Think (v2) (48.4%)

  • MMLU-ProLive

    Whether a model's knowledge holds up once guessing is made hard and questions demand multi-step reasoning rather than recall.

    TIGER-AI-Lab (Wang, Ma, Zhang et al.)3 June 2024Leader: Gemini 3.1 Pro (91.16%)

  • SimpleBenchLive

    Whether a model can handle everyday spatio-temporal reasoning, social intelligence and 'trick question' style linguistic traps that an ordinary person finds easy but that memorised knowledge doesn't help with.

    SimpleBench TeamOctober 2024Leader: Claude Fable (81.9%)

  • GPQASaturating

    Whether a model can answer graduate-level science questions that a skilled non-expert cannot solve even with unrestricted web access and half an hour per question.

    NYU, Cohere & Anthropic researchers20 November 2023Leader: Gemini 3 Pro (Deep Think) (93.8%)

  • HellaSwagSaturated

    Whether a model can pick the one commonsense-plausible continuation of an everyday situation out of several machine-generated distractors designed to fool it.

    University of Washington & Allen Institute for AI (Zellers et al.)19 May 2019Leader: GPT-4 (base, 10-shot) (95.3%)

  • AGIEvalSaturated

    Whether a model can pass the same standardised human exams — college entrance tests, law school admissions, bar exams, maths competitions — that are used to select and qualify people, rather than a bespoke academic test built only for machines.

    Microsoft Research (Zhong, Duan, Chen et al.)13 April 2023Leader: GPT-4o (71.4% (English tasks, few-shot))

  • BIG-BenchRetired

    How language-model performance and calibration change with scale across a deliberately broad, crowd-sourced set of tasks existing benchmarks did not cover — and, in the Hard subset, on the specific tasks where models still trailed average human raters.

    450+ contributors across 132 institutions (Google-led collaboration)9 June 2022

  • ARC (AI2 Reasoning Challenge)Retired

    Whether a system can answer grade-school science questions that simple retrieval and word-co-occurrence methods get wrong — a test built specifically to exclude questions that lookup alone could solve.

    Allen Institute for AI (AI2)14 March 2018

  • DROPRetired

    Whether a model can read a passage, resolve references across several parts of it, and then perform a discrete operation — such as addition, counting or sorting — on what it found, rather than lifting a single answer span directly from the text.

    AI2 & UC Irvine (Dua, Wang, Dasigi, Gardner et al.)1 March 2019Leader: GPT-4 (80.9 F1)

  • WinoGrandeRetired

    Whether a model can resolve an ambiguous pronoun using commonsense world knowledge, in sentence pairs built specifically to remove the statistical shortcuts that let earlier systems cheat at the task.

    Allen Institute for AI (Sakaguchi, Le Bras, Bhagavatula, Choi)24 July 2019

Mathematics

From grade-school word problems to unsolved research mathematics.

  • HMMTLive

    Whether a model can solve short-answer problems from the Harvard-MIT Mathematics Tournament within days of each sitting, before the problems have had time to enter training data.

    Harvard and MIT undergraduates (the underlying competition); evaluated as a live LLM benchmark by ETH Zurich's MathArena projectFebruary 2025Leader: Inkling-Small (90.2%)

  • MathArenaLive

    Can a model solve maths competition problems released after its training cutoff, so a score reflects reasoning rather than memorised answers or leaked solutions?

    ETH Zurich SRI Lab & INSAITMay 2025Leader: Claude Opus 5 (max) (84.4% ± 2.8% (overall expected performance))

  • Omni-MATHLive

    Can a model solve genuinely Olympiad-level mathematics problems, of the kind that saturated benchmarks like MATH no longer contain?

    Peking University & Alibaba10 October 2024Leader: OpenAI o1-mini (60.54%)

  • PutnamBenchLive

    Can an automated prover produce a machine-checked formal proof, not just a numeric answer, for a Putnam Competition problem?

    UT Austin (Tsoukalas, Chaudhuri et al.)15 July 2024Leader: Seed-Prover 1.5 (ByteDance) (88%)

  • HARPLive

    Whether a model's maths-competition accuracy holds up as problems get harder, using six difficulty tiers built from seven decades of US national competitions.

    Albert S. Yue, Lovish Madaan, Ted Moskovitz, DJ Strouse and Aaditya K. Singh11 December 2024Leader: o1-mini (41.1%)

  • FrontierMathDisputed

    Can a model solve original, unpublished research-level mathematics problems that resist pattern-matching against training data?

    Epoch AI8 November 2024Leader: GPT-5 (24.8% (Tiers 1-3), 8.3% (Tier 4))

  • AIMEContamination concerns

    Whether a model can solve short, single-answer competition-maths problems requiring several steps of reasoning but no written proof.

    Mathematical Association of America (the underlying competition); repurposed as an LLM benchmark by the reasoning-model research community from late 2024September 2024Leader: Inkling (97.1%)

  • MATHSaturated

    Can a model solve a competition-level mathematics problem and produce a correct step-by-step derivation, not just a lucky final number?

    UC Berkeley (Hendrycks et al.)5 March 2021

  • GSM8KSaturated

    Can a model solve a grade-school arithmetic word problem that takes several linked steps to work through, rather than a single calculation?

    OpenAI27 October 2021

  • miniF2FSaturated

    Can an automated theorem prover formally prove Olympiad-level maths statements, in a form directly comparable across different proof assistants?

    OpenAI (Zheng, Han & Polu)31 August 2021Leader: Goedel-Prover-V2-32B (90.4% (self-correction), 88.1% (standard), pass@32)

Coding & software engineering

Writing code, fixing real bugs, resolving issues in live repositories.

  • Terminal-BenchLive

    Can an AI agent actually operate a computer through a real command-line shell — issuing commands, reading their output, and adapting — to finish a multi-step task, rather than just producing plausible-looking commands?

    Stanford University & Laude Institute19 May 2025Leader: Claude Fable 5 (via Claude Code scaffold) (83.8%)

  • SWE-bench ProLive

    Can a model resolve a realistic, multi-file software engineering task in a codebase it could not have memorised — including private, commercial code rather than only well-known open-source repositories?

    Scale AI19 September 2025Leader: Muse Spark 1.1 (Meta) (61.5% ± 3.1)

  • LiveCodeBenchLive

    How well a model codes on problems it could not have memorised, by dating every problem and checking performance separately on those published before and after the model's training cutoff.

    UC Berkeley, MIT & Cornell12 March 2024Leader: DeepSeek R1-0528 (73.3%)

  • Codeforces / CodeContestsLive

    Can a model solve genuinely novel algorithmic problems under contest conditions — the kind that require devising an approach, not recalling one — well enough to rank against real competitive programmers?

    Codeforces (competitive-programming platform); CodeContests dataset curated by Google DeepMind2 February 2022Leader: Gemini 3 Deep Think (v2) (Elo 3455)

  • SWE-LancerLive

    Can a model do the paid work of a freelance software engineer — both writing code that passes real client acceptance tests, and judging which of two competing technical proposals a hiring manager should pick?

    OpenAI18 February 2025Leader: GPT-5.1-Codex-Max (79.9%)

  • Aider PolyglotLive

    Can a model act as a practical pair-programmer — reading an existing multi-language codebase, understanding a task, and editing the actual files correctly, not just writing an isolated function?

    Paul Gauthier / Aider21 December 2024Leader: GPT-5 (high reasoning effort) (88.0%)

  • BigCodeBenchLive

    Can a model write a correct program that composes multiple real library functions correctly to follow a complex, multi-step instruction — the kind of task a developer actually does, rather than an isolated algorithm puzzle?

    BigCode Project (Hugging Face, ServiceNow & collaborators)22 June 2024

  • DeepSWELive

    Can a coding agent complete an original, long-horizon software engineering task in a real repository, graded by whether the behaviour is correct — not just whether it matches one specific reference implementation?

    Datacurve8 July 2026

  • NanoGPT SpeedrunLive

    Originally, how fast a human team can train a small GPT-2-scale model to a fixed validation-loss target; adapted into an AI-agent benchmark testing whether a model can reproduce a known training-speed improvement itself, given only a hint of what changed.

    Keller Jordan & open community; formalised into an agent benchmark by Meta FAIR researchers28 May 2024

  • SWE-benchSaturating

    Can a model resolve a real, unseen GitHub issue by editing a codebase so that the project's own hidden tests pass?

    Princeton & Stanford10 October 2023Leader: Claude Opus 4.5 (state-of-the-art)

  • HumanEvalSaturated

    Can a model write a correct, working Python function from a natural-language docstring alone?

    OpenAI7 July 2021

  • MBPPSaturated

    Can a model write short, correct Python programs for simple, everyday programming tasks — the kind an entry-level programmer would be expected to solve?

    Google Research (Austin et al.)16 August 2021

  • Konwinski PrizeRetired

    Can an open-source AI system resolve real GitHub issues it could not possibly have trained on, because the test set didn't exist yet when submissions closed?

    Andy Konwinski12 December 2024

Agents, tools & computer use

Multi-step tasks: driving a browser, a terminal, a desktop, real tools.

  • OSWorldLive

    Can an agent operate a real desktop computer — a whole Ubuntu, Windows or macOS environment with real applications — to complete an open-ended task, rather than a sandboxed browser or single app?

    University of Hong Kong, Salesforce Research, Carnegie Mellon University & University of Waterloo11 April 2024Leader: Gemini 3.6 Flash (83.0%)

  • τ-benchLive

    Can an AI agent handle a realistic customer-service conversation — following a company's written policy, calling the right backend tools, and getting the outcome right — while talking to a simulated customer who has their own goals and can change their mind?

    Sierra17 June 2024Leader: Qwen3.5-397B-A17B (87.9%)

  • GAIALive

    Can an AI assistant answer real-world questions that are conceptually simple for a human but require reasoning, web browsing, tool use and handling multiple file types to actually solve?

    Meta AI (FAIR), Hugging Face & AutoGPT21 November 2023Leader: HAL Generalist Agent (Claude Sonnet 4.5) (74.55%)

  • WebArenaLive

    Can an autonomous agent operate a real, fully functional website — clicking, typing, navigating menus — to complete a task specified in natural language, the way a person actually uses the web?

    Carnegie Mellon University25 July 2023Leader: IBM CUGA (61.7%)

  • WebVoyagerLive

    Can a multimodal agent complete an everyday browsing task on a real, live website — not a sandboxed clone — by looking at the page and clicking and typing the way a person would?

    Tencent AI Lab & Westlake University25 January 2024Leader: OpenAI Operator (CUA) (87%)

  • TheAgentCompanyLive

    Can an AI agent do a full day of ordinary knowledge work inside a simulated company — writing code, filing tickets, messaging colleagues, filling in spreadsheets — well enough and consequentially enough to be judged on outcomes, not just individual isolated tasks?

    Carnegie Mellon University18 December 2024Leader: Gemini 2.5 Pro (30.3% (full completion) / 39.3% (partial credit))

  • VisualWebArenaLive

    Can a multimodal agent complete realistic web tasks that specifically require understanding an image — matching a product photo, judging a picture in a listing — not just reading and clicking text?

    Carnegie Mellon University24 January 2024Leader: Gemini 2.5 Flash (54.0%)

  • AutomationBenchLive

    Can an AI agent carry out a real business workflow across multiple SaaS applications — finding the right API endpoints itself, following a company's own layered business rules, and getting the right data to the right system — rather than completing a single, well-specified task?

    Zapier21 April 2026Leader: Claude Opus 4.7 (9.9%)

  • AndroidWorldSaturating

    Can an agent operate a real Android phone — navigating apps, typing, tapping, adjusting settings — to complete a task described in natural language, with success checked against the device's actual resulting state?

    Google Research / Google DeepMind23 May 2024Leader: AGI-0 (97.4%)

  • BrowseCompDisputed

    Can an agent find a specific, hard-to-locate fact on the open web by searching persistently and connecting scattered clues, rather than by knowing the answer already or finding it in one search?

    OpenAI10 April 2025Leader: MiniMax-M3 (83.5%)

  • Online-Mind2WebDisputed

    How well does a web agent actually perform on real, live websites under conditions close to genuine use — and how much of the field's reported progress on other web-agent benchmarks holds up under an independent, harder check?

    Ohio State University NLP group, with UC Berkeley2 April 2025Leader: OpenAI Operator (61.3%)

  • AgentBenchRetired

    How well a language model performs as an agent — not answering questions, but taking sequences of actions in interactive environments such as an operating system, a database, or a game — across eight distinct settings at once.

    Tsinghua University, with Ohio State University & UC Berkeley7 August 2023

Science & research

Domain knowledge and research work in the sciences.

  • RE-BenchLive

    How does an AI agent's performance on real, open-ended machine-learning research-engineering tasks compare with a human ML researcher's, at matched time budgets?

    METR22 November 2024

  • MLE-benchLive

    Can an AI agent do the work of a machine-learning engineer end to end — preparing data, training and tuning models, and iterating — well enough to place on a real Kaggle leaderboard?

    OpenAI9 October 2024

  • HealthBenchLive

    How well does a model handle realistic, open-ended health conversations — with a layperson or a clinician — judged against criteria that practising physicians say actually matter, rather than a multiple-choice medical exam?

    OpenAI12 May 2025Leader: OpenAI o3 (~60%)

  • PaperBenchLive

    Can an AI agent replicate a machine-learning research paper from scratch — reading it, writing the code, and running the experiments needed to reproduce its results?

    OpenAI2 April 2025

  • SciCodeLive

    Can a model write the code a scientist actually needs — numerical methods, simulations and calculations — to solve a real research problem, built up step by step rather than in one shot?

    UIUC, Argonne National Laboratory & University of Chicago (Minyang Tian et al.)18 July 2024Leader: OpenAI o3-mini-low (10.8% (main problems))

  • LAB-BenchLive

    Can a model do the practical work of biology research — finding facts in the literature, reading figures and tables, planning protocols, and reasoning over DNA and protein sequences?

    FutureHouse14 July 2024

  • ChemBenchLive

    How does a model's chemical knowledge and reasoning compare with a trained human chemist's, and does it know the limits of its own answers?

    Jablonka Lab, Friedrich Schiller University Jena1 April 2024Leader: OpenAI o1 (~92% (ChemBench-Mini))

  • SciBenchLive

    Can a model solve open-ended, college-level science problems that require multi-step quantitative reasoning, not just recall or short factual answers?

    UCLA, Caltech & University of Washington (Wang, Hu, Lu et al.)20 July 2023

Knowledge & factuality

What a model knows, and whether it admits what it does not.

  • SimpleQALive

    Whether a model gives a correct, confidently stated answer to a short factual question that has exactly one verified answer, or appropriately admits it doesn't know.

    OpenAI30 October 2024Leader: GPT-4.5 (preview) (62.5%)

  • TruthfulQALive

    Whether a model gives a truthful answer to questions written specifically to trigger common human misconceptions, rather than repeating the popular-but-false claim it likely absorbed in training.

    Oxford Future of Humanity Institute & OpenAI (Lin, Hilton & Evans)8 September 2021

  • IFEvalLive

    Whether a model reliably follows simple, mechanically checkable instructions bundled into a prompt — a word count, a required keyword, a formatting rule — rather than merely producing a plausible-sounding response.

    Google Research14 November 2023

  • FRAMESLive

    Whether a retrieval-augmented system can answer a genuinely multi-hop question that requires pulling facts from several documents and reasoning across them, not just retrieving one relevant passage.

    Google DeepMind & Harvard University19 September 2024

  • SimpleQA VerifiedLive

    The same question OpenAI's SimpleQA asks — can a model give a correct, confidently stated answer to a short factual question — on a smaller set re-checked to remove the label noise, duplication and topic imbalance found in the original.

    Google DeepMind9 September 2025Leader: Gemini 2.5 Pro (F1 55.6)

  • TriviaQASaturating

    Whether a model can answer a real trivia question, either by reading supporting evidence documents (reading comprehension) or from parametric knowledge alone (closed-book question answering).

    University of Washington (Joshi, Choi, Weld & Zettlemoyer)9 May 2017

  • MMLUSaturated

    How much a model knows across 57 academic and professional subjects, tested as four-option multiple-choice questions from elementary to expert level.

    UC Berkeley (Hendrycks et al.)7 September 2020

Language & multilingual

Understanding, translation and reasoning beyond English.

  • MMMLULive

    Whether a model retains its general-knowledge accuracy when the same MMLU questions are asked in a language other than English.

    OpenAI23 September 2024

  • MGSMLive

    Whether a model can carry out the same grade-school arithmetic word-problem reasoning in languages other than English, not just answer general-knowledge questions in translation.

    Google Research6 October 2022

  • FLORES-200Live

    How well a translation system converts text between any pair of 200 languages, using the same set of professionally translated sentences in every language.

    Meta AI (NLLB Team)6 July 2022

  • BelebeleLive

    Whether a model can read a short passage and answer a factual question about it correctly, tested in parallel across 122 languages and dialects using exactly the same underlying questions.

    Meta AI31 August 2023

  • Global-MMLULive

    Whether a model's MMLU-style general knowledge holds up across 42 languages, and separately, whether its score depends on knowledge specific to a particular culture rather than being culturally neutral.

    Cohere For AI4 December 2024

  • CMMLULive

    How much a model knows across a broad span of subjects when tested natively in Chinese, including subjects specific to Chinese culture, history and civil-service-style knowledge that an English test would not cover at all.

    MBZUAI, LibrAI, Shanghai Jiao Tong University, Microsoft Research Asia & University of Melbourne15 June 2023

  • IndQALive

    Whether a model can answer questions that require cultural and contextual knowledge specific to India, in Indian languages, rather than knowledge that happens to be translated into them.

    OpenAI5 November 2025

  • C-EvalRetired

    How much a model knows and can reason about, tested in Chinese and calibrated to the Chinese education and professional-qualification system rather than translated from an English test.

    Shanghai Jiao Tong University, Tsinghua University, University of Edinburgh & HKUST15 May 2023

Multimodal

Vision, audio, video and charts — reasoning over more than text.

  • MMMULive

    Can a model answer college-exam-level questions that genuinely require reading an accompanying image — a chart, diagram, map or chemical structure — rather than knowledge alone?

    Ohio State University & University of Waterloo (Yue, Su, Chen et al.)27 November 2023Leader: Gemini 3 Flash (81.2% (MMMU-Pro))

  • Video-MMELive

    Can a model understand a video — not just a single representative frame — across clips ranging from 11 seconds to an hour, drawing on visual content, subtitles and audio together where needed?

    Nanjing University, with XMU, HKU, PKU, CUHK, ECNU & CASIA31 May 2024Leader: Gemini 1.5 Pro (81.3% (with subtitles))

  • BLINKLive

    Whether a multimodal model can do core visual perception — judging relative depth, matching visual correspondences, spotting image tampering, reasoning across multiple viewpoints — that people solve almost instantly but that resists being reduced to language description.

    University of Pennsylvania, University of Washington, Allen Institute for AI, UC Davis and Columbia18 April 2024Leader: GPT-4o (68.0%)

  • MVBenchLive

    Whether a multimodal model can answer a question about a video that requires genuine temporal reasoning — motion, order, counting, causality — rather than being answerable from a single freeze-framed image.

    OpenGVLab, Shanghai AI Laboratory, with Nanjing University, Fudan and University of Hong Kong28 November 2023Leader: Qwen2.5-VL-72B (70.4%)

  • RealWorldQALive

    Does a model's visual understanding hold up on ordinary real-world photos — many taken from inside or around a vehicle — that test spatial reasoning and everyday physical common sense, rather than on curated benchmark imagery?

    xAIApril 2024Leader: InternVL2.5-78B (78.7%)

  • MathVistaSaturating

    Can a model do mathematical reasoning when the problem is given as a picture — a plot, a geometry diagram, a puzzle figure — rather than as text?

    UCLA, University of Washington & Microsoft Research (Lu, Bansal, Galley, Gao et al.)3 October 2023Leader: Qwen2.5-VL-72B (74.8% (testmini))

  • ScreenSpotSaturating

    Given a screenshot and a natural-language instruction, can a model click the correct on-screen element? Tests GUI grounding — locating the right icon, button or text field — across mobile, desktop and web interfaces.

    Nanjing University, with Shanghai AI Laboratory and National University of Singapore17 January 2024Leader: SeeClick (53.4%)

  • MMBenchSaturating

    How consistently, not just how often, does a vision-language model get a multiple-choice visual question right — even when the answer options are shuffled?

    OpenCompass / Shanghai AI Laboratory (Liu, Duan, Chen, Lin et al.)12 July 2023Leader: Qwen2.5-VL-72B (88.4% (MMBench-V1.1-EN test))

  • ChartQASaturating

    Can a model answer questions about a bar, line or pie chart that require reading values off it and then performing a logical or arithmetic operation on them?

    York University with NTU Singapore & Salesforce Research (Masry, Hoque, Joty et al.)19 March 2022Leader: Claude 3.5 Sonnet (90.8% (test, relaxed accuracy))

  • AI2DSaturating

    Can a model answer a multiple-choice question about a labelled science-textbook diagram — reading its text labels, following its arrows, and understanding the relationships they express — rather than just recognising objects in a photo?

    Allen Institute for AI & University of Washington24 March 2016Leader: InternVL2.5-78B (89.1% (with mask))

  • DocVQASaturated

    Can a model locate and correctly read the specific piece of text in a scanned business or government document — a form, letter, report or table — needed to answer a natural-language question about it?

    IIIT Hyderabad, CVC (Universitat Autònoma de Barcelona) & Amazon1 July 2020Leader: Qwen2.5-VL-72B (96.4% (test, ANLS))

  • VQAv2Saturated

    Given a photograph and a free-form question about it, can a model give the answer a human would — while resisting the shortcut of guessing from the question's wording alone, since every question is asked of two near-identical images with different correct answers.

    Virginia Tech, Georgia Tech and US Army Research Laboratory (Goyal et al.)2 December 2016

Safety, security & robustness

Jailbreaks, dangerous-capability evals, honesty, sabotage, harm.

  • HarmBenchLive

    How reliably a model refuses to help with a harmful request when an attacker is actively trying to jailbreak it, rather than when simply asked.

    UIUC, Center for AI Safety & collaborators6 February 2024

  • WMDPLive

    How much hazardous knowledge a model can supply in biosecurity, cybersecurity and chemical security — used both to flag dangerous capability and as a target for 'unlearning' methods that try to remove that knowledge without damaging general ability.

    Center for AI Safety & a consortium including UC Berkeley, MIT, Scale AI and SecureBio5 March 2024

  • MASKLive

    Whether a model contradicts its own stated beliefs when placed under pressure to lie — honesty, measured separately from accuracy, rather than as a proxy for it.

    Center for AI Safety & Scale AI5 March 2025Leader: Claude 3.7 Sonnet (73.4% honesty (26.6% lying rate))

  • SHADE-ArenaLive

    Whether an AI agent can secretly sabotage a task it has been trusted with while evading a monitoring system, and whether that monitor can catch it — testing agentic deception under realistic conditions rather than a single-turn chat prompt.

    Anthropic16 June 2025

  • StrongREJECTLive

    Whether a jailbroken model's response is actually useful for the forbidden request, rather than just non-refusing — correcting a pattern where earlier jailbreak evaluators counted a rambling, low-quality answer as a full success.

    UC Berkeley, Center for Human-Compatible AI15 February 2024

  • CyberSecEvalLive

    Two separate cybersecurity risks in a language model used as a coding assistant: how often it suggests insecure code, and how willing it is to help with an actual cyberattack when asked.

    Meta AI7 December 2023

  • SEC-bench ProLive

    Can a model find a genuine, previously undisclosed-style vulnerability in a large, real codebase and prove it with a working exploit input — not just patch a bug it has already been shown?

    University of Illinois Urbana-Champaign & UC Berkeley26 May 2026Leader: Codex (GPT-5.5) (58%)

  • SEC-benchLive

    Can an LLM agent handle a real software security task — reproducing a vulnerability and then patching it — in an authentic, containerised codebase?

    University of Illinois Urbana-Champaign & Purdue13 June 2025

  • AgentHarmLive

    Whether an LLM acting as an agent — using tools across multiple steps, not just chatting — will carry out a harmful task, and whether a jailbreak that fails on a chatbot still works once the model has tools to act with.

    UK AI Security Institute, Gray Swan AI & collaborators11 October 2024

  • CybenchLive

    Whether an AI agent can autonomously find and exploit real security vulnerabilities, using the same professional Capture the Flag (CTF) format security researchers train on.

    Stanford University15 August 2024

  • JailbreakBenchLive

    How well a jailbreak attack defeats a model's safety training, and how well a defence holds up against a standard set of attacks — tracked as an open, ongoing leaderboard rather than a one-off score.

    University of Pennsylvania, EPFL & collaborators28 March 2024

  • ExploitBenchLive

    How far an AI agent gets through the actual chain of an exploit — not just whether it crashes a target, but whether it can turn that crash into control of the machine.

    Carnegie Mellon University13 May 2026

  • ExploitGymLive

    Can an AI agent turn a known software vulnerability into a real, working attack — not merely identify or patch it, but exploit it end to end, including against active defences?

    Google, UC Berkeley, MPI-SP, UCSB & collaborators11 May 2026Leader: Claude Mythos Preview (157 / 898 instances)

Real-world & economic value

Professional work products and economically valuable tasks.

  • GDPvalLive

    Whether a model's output on a real occupational work task is judged, by blinded industry professionals, as good as or better than a human expert's.

    OpenAI25 September 2025Leader: GPT-5 / Claude Opus 4.1 (rated equal to or better than human work on close to half of tasks)

  • METR Time HorizonLive

    The length of a task, measured in the time a skilled human would need, that a model can complete autonomously with 50% success probability.

    METR19 March 2025Leader: Claude Opus 4.5 (~320 minutes)

  • Vending-BenchLive

    Whether an AI agent can run a simple simulated business — a vending machine — coherently over a very long horizon, rather than just complete a short task correctly.

    Andon Labs20 February 2025Leader: Claude Opus 4.7 (first place on Vending-Bench 2)

  • Agents' Last ExamLive

    Whether an AI agent can complete long-horizon, economically valuable professional tasks — not just answer questions — with a verifiable, checkable outcome.

    UC Berkeley RDI (Dawn Song et al.)3 June 2026

  • CORE-BenchLive

    Whether an AI agent can computationally reproduce the results of a published scientific paper — installing dependencies, running the authors' own code, and answering questions about the output.

    Princeton University17 September 2024Leader: CORE-Agent (GPT-4o) (45.9% overall (21% on the hardest tier))

  • WorkBenchLive

    Whether an AI agent can correctly complete a realistic office task — sending an email, scheduling a meeting, updating a record — using a sandboxed set of business tools, without taking a wrong or harmful action along the way.

    Olly Styles et al. (University of Warwick); maintained by MindsDB1 May 2024Leader: Claude Fable 5 (98% task completion, harmful action on 1.9% of tasks)

Long context & retrieval

Finding and using information across very long inputs.

  • RULERLive

    A model's 'effective context length' — the longest input it can actually use while holding accuracy above a fixed threshold — rather than the context length it claims to support.

    NVIDIA9 April 2024

  • LongBenchLive

    How well a model understands and reasons over long, realistic documents — not just whether it can retrieve one planted fact — across question answering, summarisation, few-shot learning, code and structured-data tasks.

    Tsinghua University (THUDM)28 August 2023Leader: o1-preview (57.7% (LongBench v2))

  • MRCRLive

    Whether a model can distinguish between several near-identical, repeated requests scattered through a long conversation and retrieve the correct one — a harder test than finding a single unique fact.

    Google DeepMind (introduced in the Michelangelo paper); public dataset hosted by OpenAI19 September 2024

  • InfiniteBenchSaturating

    Whether a model can process, retrieve from and reason over inputs longer than 100,000 tokens — a range existing long-context benchmarks at the time mostly stopped well short of.

    Tsinghua University (OpenBMB)21 February 2024

  • Needle in a HaystackSaturated

    Whether a model can retrieve a single fact ('the needle') planted at an arbitrary point inside a long document ('the haystack'), across varying document lengths and needle positions.

    Greg Kamradt8 November 2023

Aggregate indices & arenas

Human-preference arenas and composite indices that roll many tests into one ranking.

  • Artificial Analysis Intelligence IndexLive

    A single composite score summarising a model's capability across agentic tasks, coding, scientific reasoning and general knowledge, built as a weighted average of several independent evaluations Artificial Analysis runs itself rather than reported by the model developer.

    Artificial AnalysisQ1 2024Leader: Claude Opus 5 (max) (63)

  • LiveBenchLive

    How a model performs on recently created questions, graded by objective ground-truth answers rather than human or LLM judgment, so that scores cannot reflect memorised test data and cannot be inflated by a biased judge model.

    Abacus.AI, NYU, University of Maryland and collaborators (Colin White, Samuel Dooley, Tom Goldstein, Yann LeCun and others)24 June 2024Leader: Claude Fable 5 (Max Effort) (83.0 overall)

  • Artificial Analysis Coding Agent IndexLive

    How well a coding agent — a specific model paired with a specific harness, such as Claude Code or Codex — completes real software-engineering work end to end, not just whether the underlying model answers a coding question correctly.

    Artificial AnalysisMay 2026Leader: Claude Code – Claude Opus 5 (xhigh) / Codex – GPT-5.6 Sol (max) (67 (tied))

  • HELMLive

    How a foundation model performs not on accuracy alone, but across multiple metrics evaluated together — originally accuracy, calibration, robustness, fairness, bias, toxicity and efficiency — so that a model's trade-offs are visible rather than collapsed into one score.

    Stanford CRFM (Center for Research on Foundation Models, part of Stanford HAI)16 November 2022Leader: GPT-5 mini (2025-08-07) (0.819 mean score)

  • LMArenaDisputed

    Which of two anonymous models people prefer in open-ended, head-to-head conversation, aggregated into a running Elo rating rather than a fixed test score.

    LMSYS / UC Berkeley, later independent as LMArena (Arena Intelligence)3 May 2023Leader: Claude Fable 5 (1506 Elo)

  • Open LLM LeaderboardRetired

    How an open-weight language model scores on a fixed suite of automated academic benchmarks, run and standardised by Hugging Face rather than self-reported by the model's developer.

    Hugging FaceQ2 2023