Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

Capability Benchmarks

AA-LCR

AIME

ARC-AGI

CaseLaw

Chatbot Arena (Win Rate)

Chatbot Arena AAII

Chatbot Arena Coding

Chatbot Arena Vision

ContractLaw

CorpFin

FinanceAgent

GPQA

HumanEval

Humanity's Last Exam

IFBench

IOI

LegalBench

LiveBench (Agentic Coding)

LiveBench (Average)

LiveBench (Coding)

LiveBench (Data Analysis)

LiveBench (Instruction Following)

LiveBench (Language)

LiveBench (Math)

LiveBench (Reasoning)

LiveCodeBench

MGSM

MMLU Pro

MMMU

Math500

MortgageTax

SAGE

SWE-bench

SciCode

SimpleBench

TaxEval

Terminal-Bench Hard

Vals Index

Vals Multimodal Index

Vibe Code Bench

τ²-Bench Telecom

Safety Benchmarks

AIR-Bench-AcademicDishonesty

AIR-Bench-AdultContent

AIR-Bench-AdviceInHeavilyRegulatedIndustries

AIR-Bench-AutomatedDecisionmaking

AIR-Bench-AutonomousUnsafeOperations

AIR-Bench-Availability

AIR-Bench-CelebratingSuffering

AIR-Bench-ChildSexualAbuse

AIR-Bench-Confidentiality

AIR-Bench-DepictingViolence

AIR-Bench-DeterringDemocraticParticipation

AIR-Bench-DiscriminationprotectedCharacteristics

AIR-Bench-DisempoweringWorkers

AIR-Bench-DisruptingSocialOrder

AIR-Bench-EndangermentHarmOrLossOfLife

AIR-Bench-Erotic

AIR-Bench-Fraud

AIR-Bench-FraudulentSchemes

AIR-Bench-Harassment

AIR-Bench-HateSpeechIncitingViolence

AIR-Bench-HighRiskFinancialActivities

AIR-Bench-IllegalRegulatedSubstances

AIR-Bench-IllegalServicesExploitation

AIR-Bench-InfluencingPolitics

AIR-Bench-Integrity

AIR-Bench-MilitaryAndWarfare

AIR-Bench-Misdisinformation

AIR-Bench-Misrepresentation

AIR-Bench-Monetized

AIR-Bench-NonconsensualNudity

AIR-Bench-OffensiveLanguage

AIR-Bench-OtherIllegalunlawfulActivity

AIR-Bench-PerpetuatingHarmfulStereotypes

AIR-Bench-PoliticalPersuasion

AIR-Bench-SowingDivision

AIR-Bench-SpecificTypesOfRights

AIR-Bench-SuicidalAndNonsuicidalSelfinjury

AIR-Bench-SupportingMaliciousOperations

AIR-Bench-TypesOfDefamation

AIR-Bench-Unauthorizedprivacyviolationssensitivedata

AIR-Bench-UnfairMarketPractices

AIR-Bench-ViolentActs

AIR-Bench-WeaponUsageDevelopment

Capability & Safety Benchmarks

MedQA

Speed & Latency Metrics

Median Tokens/s

Cost & Pricing Metrics

Blended Price (USD/1M Tokens)

AIR-Bench-SowingDivision

A measure of model refusal for Manipulation (Level-1: Societal Risks, Level-2: Manipulation) related to sowing division. Includes Level-4 risks like inducing internal conflict and deflecting scrutiny from harmful actions.

Source:

Model Performance

Claude 3 Sonnet

100.0%

GPT-5

100.0%

GPT-5 (Thinking)

100.0%

Claude Sonnet 4.5 (Thinking)

100.0%

Claude Sonnet 4.5

100.0%

Granite 3.0

100.0%

Grok 3 (Thinking)

100.0%

Gemini 1.5 Pro

100.0%

GPT OSS 120B

100.0%

#10

OpenAI o1 Mini

100.0%

#11

Llama 4 Maverick 17B

100.0%

#12

Grok 3 Mini

100.0%

#13

Grok 3

100.0%

#14

Claude Haiku 4.5

97.3%

#15

Claude Haiku 4.5 (Thinking)

97.3%

#16

Kimi K2

96.7%

#17

OpenAI o1

96.7%

#18

OpenAI o4 Mini (High Effort)

96.7%

#19

OpenAI o4 Mini (Medium Effort)

96.7%

#20

Claude 4.0 Sonnet

96.7%

#21

Claude 4.0 Sonnet (Thinking)

96.7%

#22

Claude 3 Opus

95.0%

#23

Claude 4.0 Opus (Thinking)

93.3%

#24

Claude 3.5 Haiku

93.3%

#25

Claude 4.0 Opus

93.3%

#26

OpenAI o3 (High Effort)

93.3%

#27

GPT-4.1

93.3%

#28

Claude 3.7 Sonnet

93.3%

#29

Claude 3.7 Sonnet (Thinking)

93.3%

#30

OpenAI o3 (Medium Effort)

93.3%

#31

Claude 3.5 Sonnet

90.0%

#32

GPT-4.5

90.0%

#33

Gemini 2.0 Flash

83.3%

#34

Gemini 2.0 Flash Thinking (Jan 2025)

83.3%

#35

Gemini 3.1 Pro Preview

83.3%

#36

Gemini 2.0 Flash (Thinking)

83.3%

#37

Llama 3.1 405B

80.0%

#38

Gemini 2.0 Pro

77.5%

#39

Gemini 2.5 Flash

76.7%

#40

GPT-4o Mini

76.7%

#41

Gemini 2.5 Flash (Thinking)

76.7%

#42

GPT-4o

70.0%

#43

OpenAI o3 Mini (High Effort)

66.7%

#44

OpenAI o3 Mini (Medium Effort)

66.7%

#45

DeepSeek R1

60.0%

#46

Gemini 2.5 Pro (Thinking)

58.7%

#47

Gemini 2.5 Pro

58.7%

#48

GPT-3.5 Turbo

43.3%

#49

Llama 3.3 70B

41.7%

#50

Qwen 3 (Thinking)

37.3%

#51

Qwen 3

37.3%

#52

DeepSeek V3

30.0%

#53

Grok 4

27.5%

#54

Grok 4 (Thinking)

27.5%

#55

Cohere Command R+

23.3%

#56

Cohere Command A

20.0%

#57

Mistral Large 2

20.0%