Back to the leaderboard
Evaluation sources
Find what each source measures, how it updates and whether it contributes to rankings.
20Ranked sources this run69Reviewed sources & specialties
How rankings work General evaluations and user experience
17Broad capability tests and anonymous human comparisons.
- RankedAA IndexReruns current evaluations in the same environment to produce an overall capability score. Prices are reference only and do not affect the score.Artificial AnalysisEvidence budget 30%
- Cross-referenceLiveBenchContinually refreshes automatically checkable questions to see whether models remain strong on new problems.LiveBench
- RankedArena TextPeople anonymously compare answer pairs, adding an overall experience measure that multiple-choice tests miss.LMArenaEvidence budget 5%
- RankedArena WebDevAnonymous human comparisons assess whether generated websites are usable and ready to deliver.LMArenaEvidence budget 2.4%
- RankedLiveBench · Language and instructionsAn equal-weight average of Language and IF, at 50% each.LiveBenchEvidence budget 3%
- RankedArena Creative WritingAnonymous human comparisons of stories, poetry, and other creative expression assess reader engagement.LMArenaEvidence budget 5%
- ObservingDesign Arena · Visual designAnonymous human comparisons of websites, interfaces, charts, and slides.Design Arena / Intelligence
- ObservingEQ-Bench 4 · Emotional and interpersonal understandingAssesses emotions, implicit needs, and communication preferences across multi-turn conversations.EQ-Bench
- ObservingShort-Story · Short fictionIncorporates required characters, plot, and style naturally into a complete short story.Lech Mazur
- ObservingToneBench · Style and scriptsWrites video scripts in an author's specified style, assessing openings, structure, expression, and spoken pacing.Towards AI
- ObservingTubeLab · Video scriptsWrites for different video channels, comparing story organization, style, pacing, and clarity.TubeLab
- ObservingNC Bench · Author workflowsHelps authors rewrite, continue, summarize, translate, and organize creative ideas.Novelcrafter
- ObservingOpenVibeEval · Website outputsUses website outputs, accessibility, and blind preferences to examine design under different tools.OpenVibeEval
- ObservingSVGBench · Vector graphics preferencesBlind comparisons of SVGs from identical prompts assess clarity and graphic quality.SVGBench
- ObservingRapidata SVGPeople directly compare model-generated vector graphics for visual preference.Rapidata
- RankedCreative Writing v3Short creative writing and expressionEQ-BenchEvidence budget 3.6%
- RankedLongform WritingCoherence and completeness of long storiesEQ-BenchEvidence budget 2.4%
Coding
13From writing code to changing repositories: can the model build working software?
- RankedDeepSWE v1.1Changes repositories and fixes bugs in a common assistant environment to compare the underlying coding models.DataCurveEvidence budget 3.6%
- RankedTapTap MakerWrites working Lua features in a real game engine. Engine execution determines scores, without a model judge.TapTap MakerEvidence budget 3.6%
- RankedLiveBench · CodingAn equal-weight average of Coding and Agentic Coding, at 50% each.LiveBenchEvidence budget 2.4%
- ObservingHyper-τ-benchBuilds and tests working customer-service agents from business materials.Sierra
- ObservingSWE-rebenchContinually gathers real repository issues to assess repair across programming languages.Nebius
- ObservingLHTBSustained terminal use for complex engineering and tool tasks.Tencent HY / Lehigh and collaborators
- Reference onlyTerminal-Bench 4Retains original terminal-task results for models paired with different agents as cross-references.Harbor / Terminal-Bench
- Reference onlyMirrorCodeLong-running software reproductionEpoch AI
- ObservingCode MigrationSoftware migration and interface changesVals AI
- ObservingProgramBenchComplete program implementation and deliveryVals AI
- ObservingFrontierCodeQuality, tests, and scope of changes on real repository tasksCognition
- ObservingFrontierSWE v2Long-running engineering and research tasksProximal Labs
- ObservingWeirdMLImplements machine-learning solutions on unfamiliar dataWeirdML
Reasoning
11Mathematics, logic, and unfamiliar rules: can the model solve new problems?
- RankedLiveBench · Reasoning and mathematicsAn equal-weight average of Reasoning and Mathematics, at 50% each.LiveBenchEvidence budget 4.2%
- ObservingSimpleBenchEveryday common-sense and logic questions assess recognition of real constraints.SimpleBench
- RankedFrontierMath v2 · Tiers 1–3Research-level mathematical reasoningEpoch AIEvidence budget 3.6%
- RankedFrontierMath v2 · Tier 4More difficult mathematical research problemsEpoch AIEvidence budget 1.2%
- RankedChess PuzzlesFinds correct moves from textual chess positionsEpoch AIEvidence budget 1.8%
- RankedMystery Game PuzzlesDerives correct actions from unfamiliar rulesEpoch AIEvidence budget 1.2%
- ObservingMathArena · ArXivLeanVerifies mathematical proofs in LeanMathArena / ETH Zurich
- ObservingMathArena · BrokenArXivFinds and repairs errors in mathematical proofsMathArena / ETH Zurich
- ObservingMathArena · ArXivMathNew problems from recent mathematics papersMathArena / ETH Zurich
- ObservingARC-AGI-2Learns and generalizes from unfamiliar rulesARC Prize
- ObservingARC-AGI-3Learns and generalizes from unfamiliar rulesARC Prize
Knowledge
5Factual questions and graduate-level science assess knowledge and answer accuracy.
- ObservingFACTS ParametricAnswers factual questions without search to assess the model's own knowledge accuracy.Google DeepMind / Kaggle
- ObservingHumanity’s Last ExamDifficult academic knowledge and reasoning across disciplines.CAIS / Scale AI
- ObservingBullshitBench · False premisesTests whether models identify false premises rather than continue fabricating an answer.Peter Gostev / BullshitBench
- RankedSimpleQA VerifiedFactual accuracy on short questionsEpoch AIEvidence budget 4%
- RankedGPQA DiamondGraduate-level physics, chemistry, and biology knowledgeEpoch AIEvidence budget 2%
Professional work
19Financial analysis, legal advice, and banking assess completion of professional tasks.
- RankedAPEX-Agents 1.1Long investment-banking, consulting, and legal tasks assess whether a model can assist with an entire job.MercorEvidence budget 4%
- RankedVals Finance AgentEveryday financial-analyst tasks assess research reports, modeling, and related work.Vals AIEvidence budget 3%
- ObservingOfficeQA ProRetrieves information, calculates, and answers from large collections of real financial documents.Databricks
- ObservingHarvey Legal Agent BenchmarkProcesses legal materials and delivers reviewable Word, Excel, and slide files.Vals AI / Harvey
- ObservingMCP AtlasUses real MCP services to retrieve information and operate documents and databases across applications.Scale AI
- ObservingFinanceBenchmarkCompletes pricing, capital, and risk-control tasks verified by numerical answers and code execution.FinanceBenchmark
- ObservingPRBench · LegalAnswers complex legal-work questions to assess professional judgment and arguments.Scale AI
- ObservingPRBench · FinanceAnswers financial-decision questions to assess judgment, derivation, and risk explanations.Scale AI
- ObservingSpreadsheetBench 2Completes spreadsheet work including modeling, workbook repair, and visualization.SpreadsheetBench / Renmin University / AfterQuery
- ObservingVending-Bench 2Runs a simulated shop for a year, managing purchasing, negotiation, inventory, and pricing.Andon Labs
- Awaiting resultsτ³-BankingHandles banking in a common tool environment to assess rules, customer communication, and completion.SierraEvidence budget 2%
- Reference onlyEBR-benchLearns new rules and uses memory in long tasksEpoch AI
- ObservingExcel · EMBSpreadsheet editing and Excel workflowsVals AI
- ObservingLegal ResearchLegal research and argumentationVals AI
- ObservingTaxAgentBenchSpecialist tax questionsVals AI
- ObservingMedScribeOrganizing and writing clinical notesVals AI
- ObservingBioMysteryBenchBiological research reasoningVals AI
- ObservingAutomationBenchAutomates work across office applicationsZapier
- ObservingTerminal-Bench ScienceCompletes scientific tasks in a terminalHarbor
Visual understanding
2Image-understanding evidence stays in the overall leaderboard. Understanding images and creating attractive designs are different abilities.
Chinese and multilingual ability
2Supplementary language evidence. English writing is not treated as Chinese-language ability.
“Observing” means data, run conditions or usage boundaries are still being checked, and the source does not contribute to overall or category boards. Repeated collection and display slices do not add voting weight. Overlap between evaluation task sets requires continuing review.