AIEvals

Compare

Benchmarks

Choose an evaluation to inspect its task, method and published model scores.

128 of 128 benchmarksEach field uses its own task and scoring protocol
Benchmark capability, evaluation source, leading published configuration and field coverage.
BenchmarkCapabilitySourceLeading configurationScoreCoverage
SWE-Bench Pro (version unspecified)Reported by AnthropicRepository codingVendor reportsNo independent field—2 reported values
SWE-bench MultimodalReported by AnthropicRepository codingVendor reportsNo independent field—3 reported values
FrontierCode v1.1 (Extended)Reported by AnthropicAgentic codingVendor reportsNo independent field—3 reported values
ProgramBench (hidden test pass rate)Reported by AnthropicLong-context codingVendor reportsNo independent field—3 reported values
DRACO (Anthropic harness)Reported by AnthropicDeep researchVendor reportsNo independent field—4 reported values
WANDR (frozen index)Reported by AnthropicWide researchVendor reportsNo independent field—4 reported values
BenchCAD Vision2Code (no tools)Reported by AnthropicCAD reasoningVendor reportsNo independent field—5 reported values
BenchCAD Vision2Code (with tools)Reported by AnthropicCAD reasoningVendor reportsNo independent field—5 reported values
OSWorld 2.1 (offline strict pass)Reported by AnthropicComputer useVendor reportsNo independent field—6 reported values
OfficeQAReported by AnthropicDocument reasoningVendor reportsNo independent field—3 reported values
OfficeQA ProReported by AnthropicDocument reasoningVendor reportsNo independent field—3 reported values
HealthBench (length-adjusted)Reported by AnthropicHealthcareVendor reportsNo independent field—2 reported values
HealthBench (raw)Reported by AnthropicHealthcareVendor reportsNo independent field—1 reported values
HealthBench Professional (length-adjusted)Reported by AnthropicHealthcareVendor reportsNo independent field—4 reported values
HealthBench Professional (raw)Reported by AnthropicHealthcareVendor reportsNo independent field—1 reported values
PhysicianBenchReported by AnthropicHealthcareVendor reportsNo independent field—6 reported values
Global MMLUReported by AnthropicMultilingual knowledgeVendor reportsNo independent field—4 reported values
MILUReported by AnthropicMultilingual knowledgeVendor reportsNo independent field—4 reported values
SpatialBench VerifiedReported by AnthropicLife sciencesVendor reportsNo independent field—2 reported values
SingleCellBenchReported by AnthropicLife sciencesVendor reportsNo independent field—2 reported values
Morphology-to-molecule matchingReported by AnthropicLife sciencesVendor reportsNo independent field—2 reported values
Medicinal chemistry (ADME)Reported by AnthropicLife sciencesVendor reportsNo independent field—4 reported values
Protein design (sequence generation)Reported by AnthropicLife sciencesVendor reportsNo independent field—3 reported values
Protein design (library ranking)Reported by AnthropicLife sciencesVendor reportsNo independent field—3 reported values
Biomedical image analysisReported by AnthropicLife sciencesVendor reportsNo independent field—2 reported values
Protocols (troubleshooting)Reported by AnthropicLife sciencesVendor reportsNo independent field—3 reported values
Protocols (understanding V2)Reported by AnthropicLife sciencesVendor reportsNo independent field—3 reported values
APEX-Agents-AAMercor / Artificial AnalysisKnowledge workArtificial AnalysisGemini 3.5 Flash47.05%32 models
τ³-BankingSierra ResearchTool useArtificial AnalysisQwen3.8 Max51.34%154 models
ITBench-AAIBM / Artificial AnalysisIT operationsArtificial AnalysisGPT-5.6 Sol56.21%42 models
IFBenchAllen Institute for AIInstruction followingArtificial AnalysisGrok 4.20 030982.93%331 models
AIME 2025Mathematical Association of AmericaMathematicsArtificial AnalysisGPT-5.299.00%208 models
AA-AnalystAgentArtificial AnalysisData analysisArtificial AnalysisGemini 3.7 Flash60.00%33 models
MLCR-AAWisedocs / Artificial AnalysisMedical reasoningArtificial AnalysisClaude Sonnet 5.575.00%70 models
ARC-AGI-3ARC PrizeInteractive reasoningARC Prize · StandardGPT-6 Astra (Max)62.71%17 models
SWE-Bench Pro V2 FullScale AI / SEALRepository codingScale AI / SEALOpus 5 (Claude Code) xhigh99.40%10 models
SWE-Bench Pro V2 HARDScale AI / SEALRepository codingScale AI / SEALOpus 5 (Claude Code) xhigh98.00%11 models
Vals RSI IndexVals AIAutonomous researchVals AIClaude Opus 5.537.31%23 models
Vibe Code Bench 1–100Vals AIApplication developmentVals AIClaude Opus 5.530.36%20 models
CUA-benchVals AIComputer useVals AIGPT-6 Astra19.17%8 models
τ²-Bench TelecomArtificial AnalysisTool useArtificial AnalysisJT-35B-Flash99.12%323 models
Harvey LAB-AAArtificial AnalysisLegal reasoningArtificial AnalysisMuse Spark 1.395.48%42 models
EnterpriseOps-Gym-AAArtificial AnalysisEnterprise workflowsArtificial AnalysisClaude Fable 551.12%47 models
Humanity's Last Exam (text-only, without tools)Center for AI Safety and Scale AI; subset reported by DeepSeekMultidisciplinary reasoningVendor reportsNo independent field—3 reported values
Apex Bench v1.1Reported by DeepSeekReasoningVendor reportsNo independent field—6 reported values
PLawBenchPLawBench authorsProfessional knowledgeVendor reportsNo independent field—5 reported values
SWE-Bench Pro v1Scale AI / SEAL, reported by Reflection AIRepository codingVendor reportsNo independent field—5 reported values
SWE Atlas Codebase QnAScale AI / SEAL, reported by Reflection AICodebase understandingVendor reportsNo independent field—3 reported values
SWE-bench MultilingualSWE-bench, reported by Reflection AIRepository codingVendor reportsNo independent field—5 reported values
AIME 2026Mathematical Association of America, reported by Reflection AIMathematicsVendor reportsNo independent field—3 reported values
AutomationBench (public split)Zapier, reported by Reflection AIBusiness workflowsVendor reportsNo independent field—6 reported values
MCP AtlasScale AI / SEAL, reported by Reflection AIAgentic toolingVendor reportsNo independent field—7 reported values
BrowseComp (context management)OpenAI, reported by Reflection AIWeb researchVendor reportsNo independent field—4 reported values
DeepSearchQA (context management)Google DeepMind, reported by Reflection AIWeb researchVendor reportsNo independent field—2 reported values
AA-LCR (version not specified)Artificial Analysis, reported by Reflection AILong-context reasoningVendor reportsNo independent field—8 reported values
LongBench v2LongBench, reported by Reflection AILong-context reasoningVendor reportsNo independent field—4 reported values
AA-Omniscience (public split)Artificial Analysis, evaluated and reported by Reflection AIFactual knowledgeVendor reportsNo independent field—5 reported values
LABBench 2Reported by Google DeepMindScientific reasoningVendor reportsNo independent field—4 reported values
RiemannBenchSurge AI, quoted by Google DeepMindMathematicsVendor reportsNo independent field—4 reported values
GraphWalks (up to 128k)Reported by Google DeepMindLong contextVendor reportsNo independent field—4 reported values
GraphWalks (256k to 1M)Reported by Google DeepMindLong contextVendor reportsNo independent field—4 reported values
LVBenchReported by Google DeepMindVideo understandingVendor reportsNo independent field—4 reported values
CWE-bench v1CWE-bench authors, quoted by Google DeepMindSecurityVendor reportsNo independent field—4 reported values
AutomationBench-AAReported independently by Artificial AnalysisWorkflow automationArtificial AnalysisGemini 4 Argon77.51%140 models
SciCodeReported independently by Artificial AnalysisScientific codingArtificial AnalysisClaude Opus 5.566.90%144 models
GDP.pdfReported independently by Artificial AnalysisDocument reasoningArtificial AnalysisGPT-6.1 Sol31.00%130 models
CritPtReported independently by Artificial AnalysisScientific reasoningArtificial AnalysisGPT-5.6 Sol32.29%386 models
AA-OmniscienceReported independently by Artificial AnalysisFactual knowledgeArtificial AnalysisClaude Opus 5.546.42383 models
AA-LCR v1.1Reported independently by Artificial AnalysisLong-context reasoningArtificial AnalysisKimi K388.67%386 models
Terminal-Bench 4.0Terminal-Bench community and Harbor (open source)Agentic codingArtificial AnalysisClaude Sonnet 5.563.64%137 models
Terminal-Bench 2.1Terminal-Bench community and Harbor (open source)Agentic codingVals AIClaude Opus 5.587.64%75 models
SWE-bench (Verified)Princeton and Stanford researchersSoftware engineeringVals AIClaude Opus 597.00%88 models
LiveCodeBenchLiveCodeBench maintainersCode generationVals AIClaude Fable 5.190.52%143 models
GPQA DiamondRein, Hou, et al.Science reasoningVals AIGemini 3.1 Pro95.45%138 models
MMLU ProTIGER-LabKnowledgeVals AIClaude Fable 5.192.38%138 models
MMMU ProMMMU teamMultimodal reasoningVals AIClaude Fable 5.190.64%93 models
Vibe Code Bench v1.1Vals AIProduct codingVals AIClaude Sonnet 5.592.39%106 models
Terminal-Bench Science 0.1Terminal-Bench community and HarborAgentic scienceArtificial AnalysisGPT-6 Astra63.33%27 models
ProgramBenchVals AIAgentic codingVals AIClaude Opus 5.518.50%54 models
ProofBench v1.1Vals AIMathematicsVals AIClaude Sonnet 5.5100.00%46 models
LegalBenchStanford Hazy ResearchProfessional knowledgeVals AIClaude Fable 588.56%149 models
Harvey's Legal Agent BenchmarkHarvey and Vals AIProfessional knowledgeVals AIMuse Spark 1.225.42%73 models
SkillsBenchVals AIAgentic toolingVals AIDeepSeek V4.1 Flash69.80%35 models
MysteryMechanismVals AIScientific reasoningVals AIGPT-6 Astra53.15%21 models
Vals Multimodal IndexVals AICompositeVals AIClaude Fable 574.15%33 models
Vals IndexVals AICompositeVals AIGemini 4 Argon68.90%41 models
BioMysteryBenchAnthropic, run by Vals AIAgentic scienceVals AIClaude Sonnet 5.581.11%21 models
Code MigrationVals AIAgentic codingVals AIClaude Sonnet 5.569.83%71 models
CyberBench v1.1Vals AI, building on CyberGym and ARVOSecurityVals AIGPT-6 Sol77.98%43 models
EMBVals AIFinancial modellingVals AIClaude Fable 5.176.67%69 models
Finance Agent (v2)Vals AIFinancial analysisVals AIGemini 4 Argon65.40%73 models
IOIVals AI, on International Olympiad in Informatics problemsCompetitive programmingVals AIGemini 4 Argon100.00%38 models
Legal Research BenchVals AI with practising lawyersProfessional knowledgeVals AIMuse Spark 1.3 Max55.29%72 models
MedCodeVals AI with ProtegeHealthcare administrationVals AIClaude Opus 563.57%104 models
MedScribeVals AI with ProtegeHealthcare administrationVals AIClaude Opus 5.591.43%106 models
Public Benefits Bench v1.1Center for Civic Futures, run by Vals AIPublic servicesVals AIClaude Opus 576.93%45 models
SAGEVals AIEducationVals AIClaude Opus 4.756.10%90 models
SRE BenchVals AISecurityVals AIGPT-6 Astra56.87%16 models
Tax Agent BenchVals AIProfessional knowledgeVals AIClaude Fable 5.177.64%64 models
OSWorld 2.0XLANG Lab and contributors (open source)Computer useVendor reportsNo independent field—11 reported values
OSWorld 2.1XLANG Lab and contributors (open source)Computer useVendor reportsNo independent field—13 reported values
AutomationBenchZapierBusiness workflowsVendor reportsNo independent field—40 reported values
CursorBench 4.0CursorAgentic codingVendor reportsNo independent field—7 reported values
DeepSWE v1.1DatacurveSoftware engineeringDatacurveGPT-6 Astra74.12%28 models
FrontierCode v1.1Reported by OpenAI and AnthropicAgentic codingVendor reportsNo independent field—23 reported values
GDPval-AA v2.1Artificial AnalysisKnowledge workArtificial AnalysisClaude Opus 5.51,846.17188 models
AA-Briefcase v1.1Artificial AnalysisKnowledge workArtificial AnalysisClaude Opus 5.51,821.85142 models
ChartographyReported by Anthropic and Google DeepMindVisual chart recognitionVendor reportsNo independent field—39 reported values
Humanity's Last ExamCenter for AI Safety and Scale AIMultidisciplinary reasoningArtificial AnalysisClaude Opus 5.561.35%466 models
ARC-AGI-2ARC Prize FoundationAbstract reasoningVendor reportsNo independent field—5 reported values
FrontierMath Tier 4 (v2)Epoch AIMathematicsVendor reportsNo independent field—5 reported values
Agents' Last ExamReported by OpenAI and Google DeepMindComputer useVendor reportsNo independent field—28 reported values
Artificial Analysis Intelligence IndexArtificial AnalysisCompositeArtificial AnalysisClaude Opus 5.558492 models
Terminal-Bench 3.0Terminal-Bench community and Harbor (open source)Agentic codingVendor reportsNo independent field—13 reported values
NL2Repo-BenchDing et al. (2025)Agentic codingVendor reportsNo independent field—19 reported values
CyberGymSunblaze group, UC Berkeley (open source)SecurityVendor reportsNo independent field—14 reported values
SEC-bench ProSEC-bench team (open source)SecurityVendor reportsNo independent field—4 reported values
ExploitGymWang et al. (2026)SecurityVendor reportsNo independent field—20 reported values
MathArena ApexSRI Lab, ETH Zurich and INSAITMathematicsVendor reportsNo independent field—4 reported values
Codeforces (rating)Codeforces (platform); internal harness reported by DeepSeekCompetitive programmingVendor reportsNo independent field—3 reported values
BabyVisionChen et al. (2026)MultimodalVendor reportsNo independent field—9 reported values
ZeroBench (main set)Roberts et al.MultimodalVendor reportsNo independent field—9 reported values
FrontierSWEFrontierSWE community (open methodology)Agentic researchVendor reportsNo independent field—11 reported values
SWE-MarathonSWE-Marathon projectAgentic codingVendor reportsNo independent field—6 reported values
PostTrainBenchPostTrainBench authors (aisa-group), verification by Epoch AIAI researchVendor reportsNo independent field—10 reported values
ExploitBenchExploitBench (exploitbench.ai)SecurityVendor reportsNo independent field—7 reported values
Toolathlon VerifiedHKUST NLP (Toolathlon), ICLR 2026Agentic toolingVendor reportsNo independent field—8 reported values
GDPval-AA v2Reported by Z.aiKnowledge workVendor reportsNo independent field—8 reported values

Leading configurations may use different effort settings. Full methods and alternate source fields are on each benchmark page.