Seminal AI
§6

Benchmarks and leaderboards

Every model launch cites benchmark numbers, and almost every one of those numbers is less informative than it looks. A benchmark is a fixed sample of tasks with a scoring rule; a leaderboard adds a harness, a prompt, a scaffold, a sampling budget and a submitter's incentive to look good.

Data checked 2026-09-06

The gap between "model X scores 82% on SWE-bench Verified" and "model X will fix 82% of your bugs" is filled with choices — retries, test-time compute, tool access, which of six SWE-bench splits was used — that the headline number hides. This chapter covers what each major benchmark actually measures, where it has saturated, how contaminated it likely is, and who controls the scoring. All of these are free to read; running them costs inference tokens, which for agentic benchmarks can reach thousands of dollars per model.

A How to choose

Start by asking what decision the number is supposed to inform, because the benchmark families answer different questions and do not substitute for each other. For "can this model do multi-step work in my stack," only agentic, execution-verified benchmarks count — Terminal-Bench, SWE-bench Verified/Pro, tau³-bench, BFCL — and you must read the harness column, not just the score, because a 10-point gap between two models often collapses when both run under the same scaffold. For "is this model smart," the knowledge and reasoning set (GPQA Diamond, MMLU-Pro, HLE, AIME) is nearly worthless above the frontier: GPQA Diamond and AIME are saturated to the point where the remaining spread is label noise and sampling luck, and MMLU-Pro's 10-option format still rewards elimination heuristics.

Prefer benchmarks with a private or rolling held-out component — HLE's private set, ARC-AGI-2's semi-private set, MathArena's evaluate-within-days-of-the-contest policy, Epoch AI's FrontierMath — because every public static benchmark is contaminated by default and you cannot tell by how much. Trust third-party re-runs over lab-reported figures: Epoch AI and Artificial Analysis run their own harnesses and routinely land several points below vendor claims, and that delta is the single most useful signal on any launch-day chart. Treat LMArena/Arena as a measure of what people prefer reading, not what is correct — style control helps but does not remove the preference for confident, well-formatted, longer answers, and the "Leaderboard Illusion" critique about private variant testing and differential sampling has never been fully resolved.

The right default for a practitioner is: pick two or three benchmarks that resemble your workload, read the cost-per-task column next to the score, then build a 50-example internal eval on your own data, because the public number's job is to shortlist three models, not to pick one. Be actively suspicious of any benchmark where a vendor both submitted the result and controlled the scaffold, and of any score reported without a pass@k / pass^k definition.

B At a glance

Name MeasuresFormatSaturationContamination riskMaintainer Pricing
Arena (formerly LMArena / Chatbot Arena) Human pairwise preference, aggregated to a Bradley-Terry rating, across text, code, vision, document, search, image and videoLive blind A/B voting; ratings with CIs, optional Style Control adjustment; 13 category leaderboardsNever saturates by construction (relative ranking), but top-10 confidence intervals overlap heavily, making rank order largely meaninglessLow in the training-leak sense; high in the gaming sense — prompts are public and models can be tuned for voter-pleasing styleArena Intelligence (commercial spinout of LMSYS / UC Berkeley) not applicable
SWE-bench Verified Repo-level bug fixing: does the generated patch make the hidden maintainer tests pass500 human-validated instances from 12 Python repos; agentic, execution-graded, % resolved. Sibling splits: Full, Lite, Multimodal, Multilingual, Bash OnlyHeavily saturated — frontier agents report 70-80%+, and the residual is dominated by ambiguous or environment-flaky instancesVery high — public repos, public issues, public fixing commits, all pre-cutoff for every current modelSWE-bench team (Princeton/Stanford); Verified subset curated with OpenAI not applicable
SWE-bench Pro Multi-file software engineering tasks on less-contaminated and proprietary codebases1,865 tasks / 41 repos: 731 public (GPL), 276 private commercial, 858 held out; agentic, test-verified % resolvedLow to moderate — early frontier scores ~23%, since climbing past 60% on the public setLow for private and held-out splits; rising for the public GPL split now that it is publishedScale Labs (Scale AI) not applicable
GPQA Diamond Graduate-level knowledge and reasoning in biology, physics, chemistry, resistant to quick web lookup198 four-option MCQs in the Diamond subset (448 total in GPQA); single-turn, exact-match, no toolsSaturated — frontier models exceed the 65% PhD-expert baseline; remaining headroom is label noiseVery high despite gating and a canary string; widely mirrored and used in training mixesDavid Rein and co-authors; effectively unmaintained as a leaderboard, re-run by Epoch AI and Artificial Analysis not applicable
AIME / HMMT (via MathArena) Competition-level symbolic mathematics under short horizons; answer correctness, not proof quality15 integer-answer problems per AIME paper (0-999, 3 hours); MathArena runs multiple samples per model across AIME, HMMT, BRUMO, SMT, CMIMC, USAMO, IMOAIME largely saturated at the frontier; HMMT Feb and olympiad proof papers still discriminateNear-zero for a paper evaluated within days of release; near-total for any prior-year paperSRI Lab (ETH Zurich) and INSAIT for MathArena; MAA and HMMT for the underlying contests not applicable
MMLU-Pro Broad multi-domain knowledge plus reasoning across 14 academic and professional categories12,032 questions (12,000 test / 70 val), 10 answer options each, CoT prompting, exact-match accuracyApproaching saturation at the frontier; still discriminates among mid-size and open-weight modelsHigh — fully public dataset, heavily downloaded, widely present in evaluation and training mixesTIGER-Lab, University of Waterloo (last dataset update March 2026) not applicable
Humanity's Last Exam (HLE) Frontier academic knowledge and reasoning at the edge of human expertise, 100+ subjects, partly multimodal2,500 public questions (short-answer and MCQ) frozen April 2025, plus a private held-out set; accuracy and calibration errorNot saturated — Gemini 3 Pro leads around 38.3%, with GPT-5 at 25.3% and Grok 4 at 24.5%; but climbing fastModerate — public set is downloadable, mitigated by the private held-out set used to measure overfittingCenter for AI Safety (CAIS) and Scale AI; published in Nature 649:1139-1146 (Jan 2026) not applicable
ARC-AGI-2 Fluid rule induction from few examples: symbolic interpretation, compositional and context-dependent rules120 tasks each in public eval, semi-private and private sets; grid-to-grid, exact match, 2 attempts; scored against cost per task with a $10k display capFar from saturated — the stated 85% target remains open and log-linear compute scaling is documented as insufficientLow — semi-private and private sets are never published; public eval set is contaminated and used only for developmentARC Prize Foundation; ARC Prize 2026 competition on Kaggle not applicable
Terminal-Bench Agentic command-line competence: real terminal tasks graded on final system stateContainerised shell tasks run under the Harbor harness; resolution rate reported alongside cost and token usageDeliberately reset each major version — 4.0 removed saturated tasks; 2.x is largely saturated at the frontierModerate — tasks are public with a do-not-train canary string; mitigated in practice by frequent task rotationHosted by Stanford / Harbor / Laude Institute (4.0 released 28 Aug 2026) not applicable
LiveCodeBench Competitive-programming code generation, plus self-repair, test output prediction and code execution1,055 problems in v6 (May 2023-Apr 2025) from LeetCode, AtCoder and Codeforces; pass@1 / pass@5 against contest tests, filterable by release dateEasy and medium tiers saturated at the frontier; hard tier still separates modelsLow if you restrict to post-cutoff problems, which is the benchmark's core mechanism; high if you use the full setUC Berkeley LiveCodeBench team not applicable
Aider Polyglot Code editing under instruction: solution correctness plus valid, applyable edit-format compliance225 Exercism exercises in C++, Go, Java, JavaScript, Python, Rust; % correct, % correct edit format, tokens and costApproaching saturation — top reported result was GPT-5 (high) at 88.0% before updates stoppedHigh — Exercism exercises and community solutions are public and long-indexedPaul Gauthier / aider project; leaderboard last updated 20 Nov 2025 not applicable
tau-bench / tau²-bench / tau³-bench Tool-using conversational agents under domain policy: database end state, correct tool calls, reliability across repeatsSimulated user + tool APIs in retail, airline, telecom and banking-knowledge domains; pass^k reliability metric; text and full-duplex voice tracksMixed by design — τ²-bench text is near saturation (~88%), while τ³-Banking sits near 55% and remains openModerate — domains and policies are public on GitHub, though outcome-based grading limits the value of memorising transcriptsSierra (sierra.ai); τ³-bench released March 2026 not applicable
Berkeley Function Calling Leaderboard (BFCL) Function/tool calling: syntax and semantic correctness by AST match plus real execution, multi-turn, agentic, memory, web search, format sensitivityV4 category suite over enterprise and community-contributed functions; accuracy per category with an overall aggregate; reproducible via bfcl-eval pip packageSimple single-call categories are saturated; multi-turn, agentic and memory categories remain well below ceilingHigh for the static AST categories (public dataset, widely used in tool-calling fine-tuning); lower for live-execution and agentic categoriesUC Berkeley Gorilla team (Shishir Patil et al.); leaderboard updated April 2026 not applicable
MMMU / MMMU-Pro College-level multimodal understanding and reasoning over diagrams, charts, technical and scientific imagery11.5K questions, 30 subjects / 6 disciplines / 183 subfields, 30 image types; MCQ and open-ended accuracy; MMMU-Pro adds vision-only and expanded-option settingsOriginal MMMU val approaching saturation at the frontier; MMMU-Pro vision-only still leaves substantial headroomVery high since test-set answers were published in February 2026; also sourced from public textbooks and papersMMMU team, Ohio State University and University of Waterloo (Xiang Yue et al.) not applicable
Video-MME Video understanding across short, medium and long durations, with optional subtitle and audio channels900 videos / 254 hours / 2,700 human-annotated MCQ pairs; short (<2min), medium (4-15min), long (30-60min); accuracy with and without subtitlesShort subset largely saturated (frontier models ~85%); long subset still well below human performanceHigh — sourced from public YouTube video that is likely in pretraining corporaMME-Benchmarks team (github.com/BradyFU/Video-MME); dataset largely unchanged since 2024 not applicable
Artificial Analysis Composite intelligence (10 evals, weighted), plus measured output speed, TTFT, latency and price across providers and modalitiesIntelligence Index v4.2: Agents 30% (AA-Briefcase, GDPval-AA v2, τ³-Banking), General 30% (AA-Omniscience, GDP.pdf, AA-LCR v1.1), Coding 20% (Terminal-Bench v2.1, SciCode), Scientific Reasoning 20% (HLE, CritPt)Managed by rebasing — the index is re-versioned as components saturate; top models cluster around 54-57 on v4.2Inherits the contamination of its public components (HLE, Terminal-Bench, SciCode); lower for its proprietary in-house evalsArtificial Analysis (independent commercial benchmarking company) not applicable
Epoch AI Benchmarking Hub Independently re-run frontier evals: FrontierMath, SWE-bench Verified, GPQA Diamond, MATH Level 5, OTIS Mock AIME, SimpleQA Verified, MirrorCode and game-based reasoningInteractive dashboards with downloadable underlying data and documented harness configuration; multiple benchmarks per modelVaries by benchmark — GPQA Diamond and MATH L5 saturated, FrontierMath Tier 4 and Open Problems far from itLow for FrontierMath (unpublished problems, held-out set) and Open Problems; inherited-high for the public benchmarks they re-runEpoch AI (independent research organisation) not applicable
FrontierMath Research-level mathematical reasoning with computationally verifiable answers; Erdős track requires formal Lean proofsSeveral hundred unpublished problems in Tiers 1-3 and Tier 4 (v2), plus 50 Open Problems and the Lean-formalised Erdős set; accuracy, held-out subsetNot saturated — Tier 4 and Open Problems remain largely unsolved (3 of 50 Open Problems solved by AI)Very low — problems are unpublished and commissioned specifically for the benchmark, with a held-out subsetEpoch AI (creation funded by OpenAI, which drew independence criticism in 2025) not applicable

C Entries

Arena (formerly LMArena / Chatbot Arena)

Users type a prompt, get two anonymous model responses, vote for one, and the votes are fit into a Bradley-Terry (Elo-style) rating with confidence intervals. Originally LMArena out of UC Berkeley's LMSYS, it now runs as the company Arena Intelligence and splits into 13 boards: Agent, Text, WebDev, Image-to-WebDev, Vision, Document, Search, Text-to-Image, Image Edit, Text-to-Video, Image-to-Video and Video Edit. Unlike every other entry here it has no ground truth — it measures which answer a self-selected voter preferred, which is a real signal about deployment appeal and a poor one about correctness. Style Control is an offered adjustment that regresses out response length and markdown formatting; the default board and the style-controlled board disagree, and which one gets screenshotted is a rhetorical choice.

MeasuresHuman pairwise preference, aggregated to a Bradley-Terry rating, across text, code, vision, document, search, image and video
FormatLive blind A/B voting; ratings with CIs, optional Style Control adjustment; 13 category leaderboards
SaturationNever saturates by construction (relative ranking), but top-10 confidence intervals overlap heavily, making rank order largely meaningless
Contamination riskLow in the training-leak sense; high in the gaming sense — prompts are public and models can be tuned for voter-pleasing style
MaintainerArena Intelligence (commercial spinout of LMSYS / UC Berkeley)

Watch out: No ground truth: a confidently wrong, nicely formatted answer beats a correct terse one, and votes come from an unrepresentative population of AI enthusiasts. The 2025 'Leaderboard Illusion' paper alleged that large labs test many private variants and retire the losers, and that sampling rates favour incumbents — the operator disputed parts of this and the underlying practice of private pre-testing persists. Ratings for new models are unstable until vote volume accumulates, and the rebrand from lmarena.ai to arena.ai as a commercial entity introduces a conflict of interest that a university project did not have. Wrong choice when you need to know whether output is factually correct or whether code runs.

not applicable

SWE-bench Verified

An agent gets a real repository at a specific commit plus an issue description, edits the code, and the patch is scored by running the maintainers' actual test suite. Verified is the 500-instance subset that OpenAI paid annotators to filter out under-specified issues and broken tests from the original 2,294-task set, and it has become the default coding-agent number in launch posts. The site now hosts six splits — Full, Verified, Lite, Multimodal, Multilingual and Bash Only — which matter because a 'SWE-bench' claim with no split named is unreadable. Bash Only in particular constrains the agent to shell commands and typically scores well below Verified for the same model.

MeasuresRepo-level bug fixing: does the generated patch make the hidden maintainer tests pass
Format500 human-validated instances from 12 Python repos; agentic, execution-graded, % resolved. Sibling splits: Full, Lite, Multimodal, Multilingual, Bash Only
SaturationHeavily saturated — frontier agents report 70-80%+, and the residual is dominated by ambiguous or environment-flaky instances
Contamination riskVery high — public repos, public issues, public fixing commits, all pre-cutoff for every current model
MaintainerSWE-bench team (Princeton/Stanford); Verified subset curated with OpenAI

Watch out: Twelve Python repositories (django, sympy, sphinx, astropy and friends) whose entire history is in every model's training data — the issues, the discussion threads and often the exact fixing commit. Scores are dominated by the scaffold: the same model varies by double digits between a bare loop and a tuned agent, so cross-lab comparisons on the public leaderboard compare harnesses, not models. Reported figures are frequently self-submitted and unaudited, and Epoch AI's independent runs land below lab claims. The tasks are localized single-issue fixes with tests already written — nothing about design, migrations, or multi-service work. Wrong choice for anything non-Python or for greenfield code.

not applicable · open source

SWE-bench Pro

Scale's answer to SWE-bench Verified's saturation and contamination: 1,865 tasks across 41 repositories, split into 731 public instances from copyleft (GPL) open-source repos, 276 private instances taken from proprietary startup codebases, and 858 held out entirely. The GPL choice is deliberate — those repos are less likely to have been scraped into training sets — and the commercial split is genuinely unseen code. Tasks require multi-file changes rather than the single-hunk fixes that dominate Verified. The difficulty gap is the point: when top models sat around 23% on the public set they were clearing 70%+ on Verified, which is the clearest single demonstration that Verified's numbers are inflated by familiarity.

MeasuresMulti-file software engineering tasks on less-contaminated and proprietary codebases
Format1,865 tasks / 41 repos: 731 public (GPL), 276 private commercial, 858 held out; agentic, test-verified % resolved
SaturationLow to moderate — early frontier scores ~23%, since climbing past 60% on the public set
Contamination riskLow for private and held-out splits; rising for the public GPL split now that it is published
MaintainerScale Labs (Scale AI)

Watch out: The private commercial split is scored by Scale on models Scale chooses to run, which is a single-vendor bottleneck and not independently reproducible — you are trusting a company that also sells evaluation services. The public GPL split will contaminate over time exactly like Verified did. Leaderboard coverage is thinner than SWE-bench's, so the model you care about may simply be absent, and results appear in batches rather than at launch. Still Python/JS-centric despite the wider repo count.

not applicable · open source

GPQA Diamond

GPQA is 448 multiple-choice questions in biology, physics and chemistry written by PhD holders and filtered so that skilled non-experts with 30+ minutes of unrestricted web search still fail most of them — hence 'Google-proof'. Diamond is the 198-question subset where two experts agreed and the majority of non-experts failed, and it is the subset everybody quotes. It was built for scalable-oversight research, not for model ranking, and its adoption as a headline capability metric is an accident of it being the hardest thing available in 2024. The dataset is gated on Hugging Face with a canary string precisely to slow contamination, and that has plainly failed.

MeasuresGraduate-level knowledge and reasoning in biology, physics, chemistry, resistant to quick web lookup
Format198 four-option MCQs in the Diamond subset (448 total in GPQA); single-turn, exact-match, no tools
SaturationSaturated — frontier models exceed the 65% PhD-expert baseline; remaining headroom is label noise
Contamination riskVery high despite gating and a canary string; widely mirrored and used in training mixes
MaintainerDavid Rein and co-authors; effectively unmaintained as a leaderboard, re-run by Epoch AI and Artificial Analysis

Watch out: At 198 questions, one item is 0.5 percentage points, so the 1-3 point gaps labs advertise are inside the noise band before you even account for sampling temperature. Frontier models are at or above the 65% human-expert bar, so the benchmark no longer ranks the models anyone is choosing between. Contamination is effectively certain: the dataset has six-figure monthly downloads and Hugging Face lists models trained on it, and the canary string is unenforceable. Multiple-choice format rewards elimination strategies that have nothing to do with the underlying science. Wrong choice for any decision about frontier models.

not applicable · open source

AIME / HMMT (via MathArena)

AIME is the American Invitational Mathematics Examination — 15 problems, integer answers 0-999, three hours, run twice a year by the MAA — and HMMT is the Harvard-MIT Math Tournament. Neither was built as an AI benchmark; labs adopted them because they are hard, unambiguous to grade, and freshly generated each year. MathArena is the credible way to consume them: the ETH Zurich SRI Lab and INSAIT run models on each paper within days of the human contest, before solutions circulate widely, and publish standardised, multi-sample results across AIME 2025/2026, HMMT Feb and Nov, BRUMO, SMT, CMIMC, USAMO and IMO. That timing discipline is the whole value — an AIME 2024 score quoted in 2026 tells you nothing.

MeasuresCompetition-level symbolic mathematics under short horizons; answer correctness, not proof quality
Format15 integer-answer problems per AIME paper (0-999, 3 hours); MathArena runs multiple samples per model across AIME, HMMT, BRUMO, SMT, CMIMC, USAMO, IMO
SaturationAIME largely saturated at the frontier; HMMT Feb and olympiad proof papers still discriminate
Contamination riskNear-zero for a paper evaluated within days of release; near-total for any prior-year paper
MaintainerSRI Lab (ETH Zurich) and INSAIT for MathArena; MAA and HMMT for the underlying contests

Watch out: Fifteen problems per paper means each one is 6.7 percentage points; MathArena mitigates with repeated sampling but the underlying instrument is tiny and variance is large. Any AIME paper more than a few months old is thoroughly contaminated — the problems, solutions and dozens of forum threads are indexed — so labs quoting AIME 2024/2025 in 2026 are quoting a memorisation test. Scores are saturating fast on AIME while remaining lower on olympiad proof-based papers, so 'math ability' collapses to which paper you picked. Answer-only grading gives full credit for a lucky integer with a broken derivation. Wrong choice as a proxy for research mathematics — use FrontierMath for that.

not applicable · open source

MMLU-Pro

Built to fix three concrete problems with MMLU: guessing was worth 25%, many gold labels were wrong, and scores swung 4-5 points on prompt phrasing alone. MMLU-Pro raises the option count from 4 to 10 (random guessing drops to 10%), strips trivial items, adds reasoning-heavy questions, and reports prompt sensitivity down to about 2%. It covers 14 categories — math, physics, chemistry, law, engineering, economics, health, psychology, business, biology, philosophy, computer science, history and other — with 12,000 test items plus a 70-item validation split. Chain-of-thought is effectively mandatory; direct-answer evaluation drops models by roughly 19 points, so any quoted score is meaningless without the prompting mode.

MeasuresBroad multi-domain knowledge plus reasoning across 14 academic and professional categories
Format12,032 questions (12,000 test / 70 val), 10 answer options each, CoT prompting, exact-match accuracy
SaturationApproaching saturation at the frontier; still discriminates among mid-size and open-weight models
Contamination riskHigh — fully public dataset, heavily downloaded, widely present in evaluation and training mixes
MaintainerTIGER-Lab, University of Waterloo (last dataset update March 2026)

Watch out: Still multiple choice, so it measures recognition rather than generation and rewards option-elimination heuristics that have nothing to do with knowing the answer. Public and fully downloadable since 2024, with millions of downloads — treat contamination as a given. Frontier proprietary models cluster in a narrow band at the top, so it no longer separates the models most teams are actually choosing between. The dataset has been silently corrected multiple times (including 15 medical answers revised on specialist advice), which means scores from different dates are not strictly comparable. Wrong choice for anything agentic, long-context, or generative.

not applicable · open source

Humanity's Last Exam (HLE)

Nearly 1,000 subject experts from 500+ institutions in 50 countries submitted questions specifically designed to be unanswerable by then-current models; the public set was frozen at 2,500 items on 3 April 2025 and the work was published in Nature in January 2026. Crucially, CAIS and Scale keep a private held-out set of additional questions and compare public against private performance to detect overfitting — one of the few frontier benchmarks with that safeguard built in. It mixes short-answer and multiple-choice items across mathematics, humanities, natural sciences and more, and includes a multimodal fraction. An HLE-Rolling fork launched in October 2025 to accept continuing contributions as the static set ages.

MeasuresFrontier academic knowledge and reasoning at the edge of human expertise, 100+ subjects, partly multimodal
Format2,500 public questions (short-answer and MCQ) frozen April 2025, plus a private held-out set; accuracy and calibration error
SaturationNot saturated — Gemini 3 Pro leads around 38.3%, with GPT-5 at 25.3% and Grok 4 at 24.5%; but climbing fast
Contamination riskModerate — public set is downloadable, mitigated by the private held-out set used to measure overfitting
MaintainerCenter for AI Safety (CAIS) and Scale AI; published in Nature 649:1139-1146 (Jan 2026)

Watch out: Low absolute scores make it look discriminating, but the questions are deliberately adversarial trivia at the tail of human expertise — a model can score 38% on HLE and still be useless at your job, because nothing here resembles work. The published error rate in the gold answers is non-trivial, so the effective ceiling is below 100% and nobody knows exactly where. Scores are extremely sensitive to tool access and search: a browsing-enabled run is not comparable to a closed-book run, and headline numbers rarely say which was used. It is expensive to run at high reasoning effort. Wrong choice for evaluating agents, code, or long-horizon tasks.

not applicable · open source

ARC-AGI-2

Visual grid-transformation puzzles where the model infers a rule from a handful of input-output examples and applies it to a new grid. ARC-AGI-2 raises the bar over v1 with symbolic interpretation, compositional rules that interact, and rules whose application depends on context rather than surface pattern. Its calibration is the distinguishing feature: every task in every split has been solved by at least two humans in under two attempts, so a low model score is unambiguously a model failure rather than an unfair item. Three parallel 120-task sets exist — public eval, semi-private (used for verified leaderboard runs) and private (final scoring) — and the leaderboard plots score against cost per task, refusing to display runs over $10,000 total. ARC-AGI-3 has since moved to interactive agent environments, with a 2026 Kaggle competition running under strict compute limits.

MeasuresFluid rule induction from few examples: symbolic interpretation, compositional and context-dependent rules
Format120 tasks each in public eval, semi-private and private sets; grid-to-grid, exact match, 2 attempts; scored against cost per task with a $10k display cap
SaturationFar from saturated — the stated 85% target remains open and log-linear compute scaling is documented as insufficient
Contamination riskLow — semi-private and private sets are never published; public eval set is contaminated and used only for development
MaintainerARC Prize Foundation; ARC Prize 2026 competition on Kaggle

Watch out: The task format — coloured grids — is narrow and artificial, and it is genuinely unclear how much grid-puzzle skill transfers to anything a practitioner ships. Scores are dominated by test-time compute: the same model swings enormously with sampling budget, which is why the cost axis exists and why any bare percentage is unreadable. Verified leaderboard placement requires running against the semi-private set on ARC Prize's terms, so the public numbers you see are often marked 'preview' and unofficial. The 85% grand-prize threshold is far away, so for most model-selection decisions the benchmark returns 'everything is bad' and does not discriminate. Wrong choice for anything language-, code- or knowledge-related.

not applicable · open source

Terminal-Bench

Agents get a containerised shell and a task — compile something, debug a broken build, recover data, configure a service — and are graded on the resulting system state rather than on their transcript. It runs inside Harbor, a separate agent-evaluation and optimisation framework, which means the same task set can be run across many agent scaffolds and the leaderboard reports model, agent, resolution rate, cost and token count side by side. That cost column is the reason to prefer it to SWE-bench for agent selection. Versioning is aggressive: 2.0 shipped November 2025, 2.1 fixed 28 tasks in May 2026, 3.0 pushed difficulty to the frontier in July 2026, and 4.0 landed August 2026 recalibrating task resources and removing saturated tasks.

MeasuresAgentic command-line competence: real terminal tasks graded on final system state
FormatContainerised shell tasks run under the Harbor harness; resolution rate reported alongside cost and token usage
SaturationDeliberately reset each major version — 4.0 removed saturated tasks; 2.x is largely saturated at the frontier
Contamination riskModerate — tasks are public with a do-not-train canary string; mitigated in practice by frequent task rotation
MaintainerHosted by Stanford / Harbor / Laude Institute (4.0 released 28 Aug 2026)

Watch out: The version churn is a real hazard — 2.0, 2.1, 3.0 and 4.0 all shipped within ten months and scores are not comparable across them, yet vendors and even Artificial Analysis's index quote different versions concurrently. Tasks are synthetic scenarios authored for the benchmark, not sampled from real operations, so they reward a particular authored style of problem. Containerised Linux only: nothing about Windows, GUIs, or long-lived stateful production systems. Removing 'saturated tasks' in 4.0 keeps the benchmark hard but breaks longitudinal tracking by construction. Results depend as much on the Harbor agent configuration as on the model.

not applicable · open source

LiveCodeBench

The design idea is temporal: problems are scraped continuously from three competitive-programming sites and tagged with a publication date, so you can evaluate any model only on problems that appeared after its training cutoff. v5 (January 2025) held 880 problems spanning May 2023 to January 2025; v6 (April 2025) expanded to 1,055 through April 2025. Beyond plain code generation it scores three other scenarios — self-repair, test output prediction and code execution — which separate 'can write a solution' from 'can predict what code does'. Grading is pass@1 / pass@5 against the contest test cases, so it is fully execution-verified with no judge model.

MeasuresCompetitive-programming code generation, plus self-repair, test output prediction and code execution
Format1,055 problems in v6 (May 2023-Apr 2025) from LeetCode, AtCoder and Codeforces; pass@1 / pass@5 against contest tests, filterable by release date
SaturationEasy and medium tiers saturated at the frontier; hard tier still separates models
Contamination riskLow if you restrict to post-cutoff problems, which is the benchmark's core mechanism; high if you use the full set
MaintainerUC Berkeley LiveCodeBench team

Watch out: Competitive programming is not software engineering — a model that tops this can still be poor at reading an unfamiliar codebase, and vendors sometimes cite it as a proxy for engineering ability, which it is not. The date-filter defence only works if you know the true cutoff and if the problem was not discussed on forums before the stated date; both assumptions leak. Windows are short (v6 ended April 2025), so for a 2026 model the uncontaminated slice is small and the sub-sample noise is correspondingly large. Different published scores use different date windows and different versions, making cross-report comparison unreliable unless the window is stated. Python-dominant.

not applicable · open source

Aider Polyglot

Distinct from other coding benchmarks because it grades two things at once: whether the solution is right, and whether the model produced a syntactically valid edit that the tool could apply without human intervention. It runs 225 of the hardest Exercism exercises across C++, Go, Java, JavaScript, Python and Rust, reporting pass rate, edit-format compliance, token usage and dollar cost per run. That edit-format column is the unique signal — a model that solves the problem but mangles its search/replace blocks is useless inside a coding tool, and no other public benchmark isolates that. Note the leaderboard's last update was 20 November 2025.

MeasuresCode editing under instruction: solution correctness plus valid, applyable edit-format compliance
Format225 Exercism exercises in C++, Go, Java, JavaScript, Python, Rust; % correct, % correct edit format, tokens and cost
SaturationApproaching saturation — top reported result was GPT-5 (high) at 88.0% before updates stopped
Contamination riskHigh — Exercism exercises and community solutions are public and long-indexed
MaintainerPaul Gauthier / aider project; leaderboard last updated 20 Nov 2025

Watch out: The leaderboard has not been refreshed since November 2025, so it is roughly ten months stale as of September 2026 and omits every model released since — treat it as a historical reference, not a current shortlist. Exercism exercises are self-contained puzzles with existing tests, nothing like navigating a real repository. Results are partly a measure of how well a model matches aider's specific diff format, which does not generalise to tools that use other edit protocols. Small sample (225 items) and single-maintainer governance mean it can go quiet without notice. Wrong choice for agentic or multi-file work.

not applicable · open source

tau-bench / tau²-bench / tau³-bench

An agent handles a simulated customer (played by an LLM) in a domain with real tools and a real database, and is scored on whether the database ends in the correct state and the right tool calls were made — not on whether the transcript reads nicely. Its signature metric is pass^k: the probability that the agent succeeds on all k independent attempts at the same task, which punishes flakiness in a way pass@k explicitly rewards. The line has evolved through three generations: τ-bench (June 2024, retail and airline), τ²-bench (June 2025, adds telecom and a dual-control setting where the user also acts), and τ³-bench (March 2026), which adds a banking knowledge domain with ~700 documents and RAG retrieval, plus full-duplex real-time voice evaluation with interruptions, accents and background noise.

MeasuresTool-using conversational agents under domain policy: database end state, correct tool calls, reliability across repeats
FormatSimulated user + tool APIs in retail, airline, telecom and banking-knowledge domains; pass^k reliability metric; text and full-duplex voice tracks
SaturationMixed by design — τ²-bench text is near saturation (~88%), while τ³-Banking sits near 55% and remains open
Contamination riskModerate — domains and policies are public on GitHub, though outcome-based grading limits the value of memorising transcripts
MaintainerSierra (sierra.ai); τ³-bench released March 2026

Watch out: The simulated user is itself an LLM, so a chunk of the variance is the simulator's behaviour rather than the agent's — change the user model and the scores move. Domains are small and hand-authored (retail, airline, telecom, banking), and the policies are Sierra's inventions, so a high score means 'follows these fictional policies', not 'follows yours'. Sierra sells customer-service agents, which is a direct commercial interest in the benchmark that defines the category. Task-level bugs have been material: the v1.0.x releases shipped 75+ task fixes for wrong actions, ambiguous instructions and impossible constraints, so older published scores are not comparable. Voice evaluation adds provider-specific latency effects that confound model quality.

not applicable · open source

Berkeley Function Calling Leaderboard (BFCL)

The reference benchmark for whether a model emits well-formed, correct tool calls. It grades two ways: abstract-syntax-tree comparison against a reference call (checking function name, parameter names and value types without needing to execute), and real execution against live APIs. V4 extends beyond single calls into holistic agentic evaluation, adding web-search tasks, memory, multi-turn interaction, enterprise and community-contributed function sets, and a format-sensitivity category that measures how much a prompt-based model's tool-calling degrades when the schema is presented differently. Everything is reproducible: results are pinned to a commit, and a versioned pip package (bfcl-eval) lets you re-run the exact harness yourself.

MeasuresFunction/tool calling: syntax and semantic correctness by AST match plus real execution, multi-turn, agentic, memory, web search, format sensitivity
FormatV4 category suite over enterprise and community-contributed functions; accuracy per category with an overall aggregate; reproducible via bfcl-eval pip package
SaturationSimple single-call categories are saturated; multi-turn, agentic and memory categories remain well below ceiling
Contamination riskHigh for the static AST categories (public dataset, widely used in tool-calling fine-tuning); lower for live-execution and agentic categories
MaintainerUC Berkeley Gorilla team (Shishir Patil et al.); leaderboard updated April 2026

Watch out: AST matching accepts only calls that resemble the reference, so a legitimately different-but-correct call can be scored wrong, and it says nothing about whether calling that function was a good idea. Live-execution categories depend on third-party APIs that change and break, making historical scores non-reproducible in practice despite the pinned commits. The functions are toy or synthetic relative to a real enterprise tool surface with dozens of overlapping endpoints. Native tool-calling models and prompt-based models are evaluated under different conditions, so cross-category comparison is misleading. Leaderboard refreshes are periodic rather than at-launch, so brand-new models may be missing.

not applicable · open source

MMMU / MMMU-Pro

The default multimodal knowledge benchmark: 11,500 college-exam questions across 30 subjects in 6 disciplines and 183 subfields, paired with 30 heterogeneous image types — diagrams, charts, chemical structures, medical images, music scores, circuits, maps. MMMU-Pro (September 2024) is the hardened variant that filters out questions answerable from text alone, expands the candidate options, and adds a vision-only setting where the question itself is embedded in a screenshot, forcing the model to actually read the image rather than shortcut through the caption. That text-only-solvable filtering is important, because a meaningful fraction of original MMMU questions could be answered with no image at all.

MeasuresCollege-level multimodal understanding and reasoning over diagrams, charts, technical and scientific imagery
Format11.5K questions, 30 subjects / 6 disciplines / 183 subfields, 30 image types; MCQ and open-ended accuracy; MMMU-Pro adds vision-only and expanded-option settings
SaturationOriginal MMMU val approaching saturation at the frontier; MMMU-Pro vision-only still leaves substantial headroom
Contamination riskVery high since test-set answers were published in February 2026; also sourced from public textbooks and papers
MaintainerMMMU team, Ohio State University and University of Waterloo (Xiang Yue et al.)

Watch out: The test-set answers were released publicly on 12 February 2026 to allow local evaluation, which destroys the held-out property — any test-split score published after that date should be treated as self-reported and contamination-exposed. A slice of original MMMU is solvable without looking at the image; use MMMU-Pro if you care about that, and note the two are frequently conflated in launch charts. It is an exam benchmark, so it measures academic recall with pictures, not document processing, OCR quality, or visual grounding for agents. Image quality is uneven since figures were sourced from textbooks and papers. Wrong choice for evaluating screenshot-driven computer-use agents.

not applicable · open source

Video-MME

The most-cited general video understanding benchmark. It splits 900 videos into short (under 2 minutes), medium (4-15 minutes) and long (30-60 minutes) buckets so you can see exactly where a model's temporal reasoning falls apart — the long bucket is where most video models collapse and where the useful signal lives. All 2,700 QA pairs are human-annotated rather than generated. It supports separate evaluation with and without subtitles (744 subtitle tracks provided) and with audio, which is diagnostically important: a large with-subtitles gain means the model is reading text, not watching video.

MeasuresVideo understanding across short, medium and long durations, with optional subtitle and audio channels
Format900 videos / 254 hours / 2,700 human-annotated MCQ pairs; short (<2min), medium (4-15min), long (30-60min); accuracy with and without subtitles
SaturationShort subset largely saturated (frontier models ~85%); long subset still well below human performance
Contamination riskHigh — sourced from public YouTube video that is likely in pretraining corpora
MaintainerMME-Benchmarks team (github.com/BradyFU/Video-MME); dataset largely unchanged since 2024

Watch out: The academic-only licence is a genuine blocker for commercial evaluation — you cannot legally use it to qualify a vendor for a product, unlike almost everything else in this chapter. Multiple choice, so it rewards elimination and does not test description, grounding or temporal localisation. Frame-sampling strategy dominates results: the same model scores very differently at 8 vs 64 vs 256 frames, and published numbers rarely state the frame budget, which makes cross-report comparison close to meaningless. YouTube-sourced videos are inherently contaminated for any model trained on public web video. The benchmark has not been substantially revised since 2024 while video models have advanced considerably.

not applicable · open source

Artificial Analysis

The most-used meta-tracker. It maintains internal copies of every evaluation dataset and runs them on its own infrastructure, so its numbers are not vendor self-reports — the deltas against launch-post claims are frequently the most useful thing on the page. Intelligence Index v4.2 aggregates ten evaluations in four weighted groups: Agents at 30% (AA-Briefcase, GDPval-AA v2, τ³-Banking), General at 30% (AA-Omniscience, GDP.pdf, AA-LCR v1.1), Coding at 20% (Terminal-Bench v2.1, SciCode) and Scientific Reasoning at 20% (Humanity's Last Exam, CritPt). Alongside intelligence it publishes measured throughput, time-to-first-token, end-to-end latency under load and per-token pricing across serving providers, and it extends to speech, image, video, music and coding agents. It also sells commercial products (Optima, MicroEvals).

MeasuresComposite intelligence (10 evals, weighted), plus measured output speed, TTFT, latency and price across providers and modalities
FormatIntelligence Index v4.2: Agents 30% (AA-Briefcase, GDPval-AA v2, τ³-Banking), General 30% (AA-Omniscience, GDP.pdf, AA-LCR v1.1), Coding 20% (Terminal-Bench v2.1, SciCode), Scientific Reasoning 20% (HLE, CritPt)
SaturationManaged by rebasing — the index is re-versioned as components saturate; top models cluster around 54-57 on v4.2
Contamination riskInherits the contamination of its public components (HLE, Terminal-Bench, SciCode); lower for its proprietary in-house evals
MaintainerArtificial Analysis (independent commercial benchmarking company)

Watch out: A single weighted composite hides everything that matters: two models can tie at 54 with completely different profiles, and the 30/30/20/20 weighting is Artificial Analysis's editorial judgement, not a fact about intelligence. The index rebases with each version, so v4.2 scores are not comparable to earlier versions and historical charts quietly break. Component versions lag — v4.2 uses Terminal-Bench v2.1 while the benchmark itself is on 4.0 — so a component can be saturated or superseded while still carrying 20% weight. Half the components are proprietary AA-built evals (AA-Briefcase, AA-Omniscience, AA-LCR) that nobody can independently reproduce or audit. Published cost figures exclude judge calls and preprocessing, so the real cost of a run is higher than shown. It is a commercial company selling evaluation products to the same industry it ranks.

not applicable

Epoch AI Benchmarking Hub

Epoch AI runs an independent benchmarking hub covering FrontierMath (Tiers 1-3 v2, Tier 4 v2, and the Lean-formalised Erdős set), SWE-bench Verified, GPQA Diamond, MATH Level 5, OTIS Mock AIME 2024-2025, SimpleQA Verified, MirrorCode, chess puzzles, Earthborne Rangers (EBR-bench) and mystery-game puzzles. The value is methodological transparency rather than novelty: they publish harness details, per-model run configurations and the raw result data alongside the leaderboards, which makes their numbers auditable in a way vendor blog posts are not. They also maintain the wider Epoch datasets on training compute, model releases and hardware trends, so a benchmark result sits next to the compute that produced it. The FrontierMath Open Problems expansion adds 50 genuinely unsolved research questions, of which AI has solved three so far.

MeasuresIndependently re-run frontier evals: FrontierMath, SWE-bench Verified, GPQA Diamond, MATH Level 5, OTIS Mock AIME, SimpleQA Verified, MirrorCode and game-based reasoning
FormatInteractive dashboards with downloadable underlying data and documented harness configuration; multiple benchmarks per model
SaturationVaries by benchmark — GPQA Diamond and MATH L5 saturated, FrontierMath Tier 4 and Open Problems far from it
Contamination riskLow for FrontierMath (unpublished problems, held-out set) and Open Problems; inherited-high for the public benchmarks they re-run
MaintainerEpoch AI (independent research organisation)

Watch out: Coverage is deliberately narrow and updates lag launches by days to weeks, so it is a verification layer rather than a first look. Because they use their own harness and prompting, their figures are systematically lower than vendor claims — correct, but it means you cannot mix an Epoch number and a lab number in the same chart. FrontierMath's OpenAI funding, and OpenAI's access to problem data, drew sustained criticism in 2025 over whether an independent benchmark can be commissioned by an evaluated party; disclosure improved but the structural conflict is real. Several of their benchmarks (chess puzzles, board-game and mystery-game evals) are research instruments with little bearing on production model selection. Small team, so continuity depends on funding.

not applicable · open source

FrontierMath

Commissioned from professional mathematicians and never published, which makes it one of the few genuinely uncontaminated hard-reasoning benchmarks available. Problems span Tiers 1-3 (undergraduate through graduate difficulty) and Tier 4 (research level), and every problem has a computationally verifiable answer so grading needs no human and no LLM judge. Two extensions matter: Open Problems is 50 significant unsolved research questions where AI has so far cracked three, and FrontierMath Erdős formalises problems studied by Paul Erdős in the Lean proof assistant, requiring a complete machine-checked proof or disproof rather than an answer. That Lean track is the strongest available evidence about whether a model can do mathematics rather than pattern-match to it.

MeasuresResearch-level mathematical reasoning with computationally verifiable answers; Erdős track requires formal Lean proofs
FormatSeveral hundred unpublished problems in Tiers 1-3 and Tier 4 (v2), plus 50 Open Problems and the Lean-formalised Erdős set; accuracy, held-out subset
SaturationNot saturated — Tier 4 and Open Problems remain largely unsolved (3 of 50 Open Problems solved by AI)
Contamination riskVery low — problems are unpublished and commissioned specifically for the benchmark, with a held-out subset
MaintainerEpoch AI (creation funded by OpenAI, which drew independence criticism in 2025)

Watch out: You cannot run it yourself — the problems are held by Epoch, so every number is Epoch's number and there is no way to reproduce, audit the items, or evaluate your own fine-tune. OpenAI funded the benchmark's creation and had access to problem data, which triggered a substantial 2025 controversy about independence; Epoch improved disclosure and maintains a held-out set, but a funder-evaluated-by-the-funded arrangement remains structurally compromised. Answer-verifiable problems are a biased sample of mathematics — they exclude most of what research mathematicians actually do, which is why the Lean Erdős track exists. Absolute scores are low and move in lumps, so it discriminates poorly among mid-tier models. Wrong choice for any practical model-selection decision.

not applicable