Claude Opus 5

Claude Opus 5 is a reasoning model from Anthropic. 72 benchmarks count toward its index, in 8 categories.

Closed weights
Reasoning
Input: text · image · file
Released 2026-07

Ranked

Index73.8±1.9Median 51.0

Coverage100%

Speed48/sMedian 67/s

First token0.9 sMedian 1.3 s

Input / 1M$5.00batch $2.50Median $0.60

Output / 1M$25.00batch $12.50 Median $2.55

Context1MMedian 262K

Elo1502±4 #7Median 1422

Details

From OpenRouter

API ID: anthropic/claude-opus-5
Max output: 128K
Knowledge cutoff: —
Tools: Yes
JSON output: Yes
Cache read / 1M: $0.50
Weights: —
Retires: —

Where to run

7 providers

ProviderTypeInputOutputContext

Amazon BedrockCloud$5.00$25.001M

AnthropicMaker$5.00$25.001M

DigitalOceanCloud$5.00$25.001M

Google Vertex AICloud$5.00$25.001M

Microsoft AzureCloud$5.00$25.001M

OpenRouterRouter$5.00$25.001M

Snowflake CortexCloud$5.00$25.001M

US dollars per 1M tokens, list price. From models.dev (opens in a new tab) (MIT).

Capabilities

Out of 100

Agentic23 counted71.620 / 218

Coding23 counted74.810 / 209

Reasoning10 counted75.416 / 227

Multimodal2 counted72.56 / 84

Knowledge7 counted73.110 / 268

Multilingual1 counted63.6—

Instruction1 counted55.6—

Math5 counted74.815 / 168

Results

72 counted

BenchmarkResultPlace

Software Engineering Benchmark VerifiedCoding96.0%1 / 114

SWE-benchCoding97.0% ±0.81 / 87

MMLU ProKnowledge91.6% ±0.32 / 135

IOI v1Coding91.7% ±8.31 / 61

LiveBench ResearchAgentic88.9%1 / 52

MMMU ProMultimodal89.9% ±0.72 / 92

LiveCodeBenchCoding89.0% ±0.94 / 142

Artificial Analysis Humanity's Last ExamKnowledge54.9%7 / 222

ARC-AGI-1 (semi-private)Reasoning97.5%5 / 126

Agent Arena command recoveryAgentic13.62 / 50

Artificial Analysis Omniscience AccuracyKnowledge60.9%7 / 164

Toolathlon-VerifiedAgentic80.6%1 / 23

GPQA Diamond · Vals AIKnowledge93.4% ±1.27 / 135

Humanity's Last ExamKnowledge64.7%4 / 77

ProgramBench: Can Language Models Rebuild Programs From Scratch?Coding93.0%1 / 19

GDPval-AA normalizedAgentic61.1%7 / 133

Artificial Analysis MMMU-ProMultimodal84.7%6 / 113

Artificial Analysis GPQA DiamondKnowledge93.2%11 / 199

MCP AtlasAgentic85.8%3 / 52

LiveBench Agentic CodingAgentic65.2%4 / 66

Critical Physics TasksReasoningAbove fit29.1%13 / 214

ARC-AGI-2 (semi-private)Reasoning90.4%8 / 131

GPQA diamond · Epoch AIKnowledge93.9% ±1.514 / 217

Terminal-Bench 2.1 · Vals AIAgentic84.6% ±1.05 / 75

LiveBench LanguageKnowledge88.7%5 / 66

Terminal-Bench 3.0Agentic42.7%1 / 13

DeepSearchQAAgentic95.0%2 / 24

INCLUDEMultilingual89.8%1 / 12

Humanity's Last Exam with toolsAgentic64.7%3 / 36

SWE-bench ProCoding79.2%7 / 83

Vibe Code Bench v1.1Coding88.4% ±3.011 / 123

Artificial Analysis Terminal-Bench v2.1Agentic89.1%2 / 21

LiveBench ReasoningReasoning91.2%8 / 80

MirrorCodeCoding87.8% ±7.24 / 39

Code MigrationCoding57.5% ±4.48 / 77

τ²-bench BankingAgentic48.7%3 / 28

OTIS Mock AIME 2024-2025Math98.9% ±1.122 / 200

ProofBench v1.1Math99.0% ±1.06 / 54

Mystery Game PuzzlesReasoning59.0% ±4.99 / 80

Humanity's Last Exam without toolsKnowledge56.3%6 / 53

LiveBench DesignAgentic83.0%6 / 52

Vibe Code Bench 1-100Coding28.5% ±4.33 / 26

OSWorld 2.0Agentic70.6%3 / 25

BrowseCompAgentic90.8%8 / 63

LiveBench Long AgenticAgentic62.1%7 / 51

MazeBench: visual spatial reasoning in a 3D open worldAgentic12.0%7 / 48

Terminal-Bench 4.0Agentic53.5% ±1.310 / 61

FrontierMath-Tiers-1-3-v2-PrivateMath85.6% ±2.115 / 90

ApprenticeBench: end-to-end computer use, continual learning, and long-horizon agency on a real accounts-payable jobAgentic36.0%3 / 18

FrontierCode 1.1 MainCoding53.4%3 / 17

IOICoding84.3% ±10.08 / 45

LiveBench MathematicsMath95.7%12 / 66

Agent Arena task outcomeAgentic10.210 / 54

Artificial Analysis AnalystAgentAgentic53.8%5 / 26

LiveBench CodingCoding81.4%13 / 66

SkillsBenchCoding60.4% ±4.67 / 35

FrontierMath-Tier-4-v2-PrivateMath73.2% ±7.015 / 72

SimpleQA VerifiedKnowledge59.9% ±1.619 / 91

LiveBench Coding 2.0Coding90.8%11 / 52

cursorBench32Coding70.0%4 / 18

LiveBench Browser UseAgentic84.0%11 / 49

Artificial Analysis SciCodeCoding56.4%26 / 114

ARC-AGI-3 (semi-private)Reasoning30.2%6 / 26

Artificial Analysis Long Context ReasoningReasoning79.3%52 / 217

Terminal-Bench 4.1.0Agentic52.1% ±1.76 / 23

Terminal-Bench 4.0.0Agentic51.8% ±1.76 / 23

Furniture AssemblyReasoning60.8% ±6.210 / 37

ScarfBench: Cross-Framework Application Migration in Enterprise JavaCoding69.8%3 / 11

VulcanBench v3Coding87.0%4 / 14

OfficeQA ProMultimodal66.9%5 / 16

Agent Arena steerabilityAgentic4.417 / 54

EBR-benchReasoning45.7% ±3.610 / 31

Chess PuzzlesReasoning43.0% ±5.031 / 90

FrontierSWE v2Coding52.0%8 / 22

Artificial Analysis DeepsecBench-AAAgentic25.9%8 / 21

DeepSWEAgentic68.8%24 / 59

PostTrainBench v1.1Coding35.0%7 / 15

cursorBench40Coding46.6%10 / 21

FrontierCode 2.0Coding26.7%13 / 27

LiveBench Data AnalysisReasoning74.6%40 / 80

Bug Hunt BenchCoding27.0%15 / 25

HealthBench ProfessionalKnowledge59.8%12 / 18

LiveBench Instruction FollowingInstruction63.8%46 / 66

OSWorld 2.1 (all 108 tasks)Agentic74.0%8 / 11

AutomationBenchAgentic26.0%31 / 32

CorpFin v2Finance73.2% ±0.91 / 131

MedCodeHealthcare63.6% ±2.01 / 106

MortgageTaxFinance72.1% ±0.91 / 97

ARC-AGI-1 (public eval)Reasoning99.0%2 / 116

Public Benefits Bench v1.1Public benefits76.9% ±1.11 / 50

Legal Research BenchLegal55.3% ±3.52 / 90

MedScribeHealthcare91.0% ±1.94 / 109

LegalBenchLegal87.0% ±0.47 / 149

ARC-AGI-2 (public eval)Reasoning97.1%6 / 122

Tax Agent BenchFinance47.1% ±3.65 / 82

EMBFinance73.6% ±2.26 / 75

ProgramBenchCoding3.0% ±1.26 / 65

Finance Agent (v2)Finance58.6% ±0.110 / 79

BioMysteryBenchScience79.3% ±2.14 / 30

TaxEval v2Finance75.1% ±0.823 / 143

Terminal-Bench-ScienceScience27.1% ±5.410 / 44

MysteryMechanismScience37.4% ±3.38 / 30

SRE BenchOther12.2% ±2.011 / 36

CUA-benchOther9.0%4 / 9

Time Horizon Index: KSPOther18.8%6 / 13

FrontierMath-ErdosMath0.0%8 / 15

CyberBench v1.1Other65.4% ±5.532 / 48

VulcanBench Coding Intelligence Index v1Coding96.4%—

Legal Agent Benchmark mean criterion-pass rate — Harvey held-out setAgentic94.1%—

Legal Agent Benchmark mean criterion-pass rate — Anthropic harnessAgentic93.7%—

Multi-Agent BrowseComp — 10-agent team prerelease configurationAgentic93.6%—

Artificial Analysis Harvey LAB-AA v1.0Agentic93.5%—

Global MMLUMultilingual92.5%—

Multi-task Indic Language Understanding BenchmarkMultilingual92.1%—

ArXivMath June 2026 with toolsMath91.3%—

ArXivMath June 2026 without toolsMath90.8%—

BioMysteryBench Human SolvableKnowledge90.1%—

MCP-Atlas mean claim coverageAgentic89.1%—

Data Research and Analysis with Complex OperationsAgentic88.6%—

Toolathlon Verified Pass@3Agentic87.0%—

GDP.pdf mean criteria pass rate with toolsMultimodal85.5%—

LABBench2: An Improved Benchmark for AI Systems Performing Biology ResearchKnowledge84.2%—

GDP.pdf mean criteria pass rate without toolsMultimodal83.4%—

ProgramBench hidden-test pass rate after episode 1Coding83.0%—

Chartography with image and code toolsMultimodal83.0%—

BenchCAD Vision2Code voxel IoU with toolsMultimodal82.1%—

RiemannBench with toolsMath79.0%—

Benchling Molecular Biology Protocols UnderstandingKnowledge78.4%—

OfficeQAMultimodal78.1%—

HealthBench Professional raw scoreKnowledge73.4%—

Toolathlon Verified Pass cubedAgentic73.1%—

LatchBio SpatialBench VerifiedKnowledge72.5%—

HealthBench raw scoreKnowledge67.1%—

FrontierCode 1.1 ExtendedCoding63.6%—

Anthropic Organic Chemistry V2 evaluationKnowledge61.6%—

Molecular Biology Protocols TroubleshootingKnowledge61.1%—

LatchBio SingleCellBenchKnowledge60.6%—

RiemannBench without toolsMath60.0%—

HealthBench length-adjusted scoreKnowledge57.8%—

HLE-VerifiedKnowledge54.4%—

BioMysteryBench Human DifficultKnowledge49.4%—

ProteinGym HardKnowledge47.7%—

Artificial Analysis EnterpriseOps-GymAgentic47.5%—

Anthropic Protein Design evaluationKnowledge42.5%—

International Mathematical Olympiad 2026Math42.0%—

BenchCAD Vision2Code voxel IoU without toolsMultimodal36.6%—

Chartography without toolsMultimodal29.6%—

Legal Agent Benchmark all-pass rate — Anthropic harnessAgentic23.6%—

Toolathlon Verified average assistant turnsAgentic23.5%—

Legal Agent Benchmark all-pass rate — Harvey held-out setAgentic11.7%—

Sources

BenchLMCC BY-NC 4.0 OpenRouterTerms restrict reuse Vals AINo licence stated Epoch AICC BY ARC Prize FoundationTerms restrict reuse LiveBenchNo licence stated Terminal-BenchNo licence stated Sierra ResearchMIT LMArenaCC BY 4.0 models.devMIT

History

5 releases

Index75.4 → 73.8

23 Oct 202612 Nov 2026

Output / 1M$25.00 → $25.00

23 Oct 202612 Nov 2026

Questions

How much does Claude Opus 5 cost?: $5.00 per 1M input tokens, $25.00 per 1M output tokens.
How good is Claude Opus 5?: Its index is 73.8 out of 100. That is place 14 of 283 ranked models.
What is Claude Opus 5 best at?: Knowledge: 73.1 out of 100, place 10 of 268.
How fast is Claude Opus 5?: 48 output tokens per second, 0.9 s to the first token.
What is the context window of Claude Opus 5?: 1M tokens.
Is Claude Opus 5 open source?: No. Its weights are not published.
When was Claude Opus 5 released?: July 2026.

Badge

HTML: <a href="https://inference360.ai/models/claude-opus-5"><img src="https://inference360.ai/badges/claude-opus-5.svg" alt="Claude Opus 5 on Inference360: place 14 of 283, index 73.8" height="20"></a>
Markdown: [![Claude Opus 5 on Inference360: place 14 of 283, index 73.8](https://inference360.ai/badges/claude-opus-5.svg)](https://inference360.ai/models/claude-opus-5)

Compare with

Claude Fable 574.0 GPT-5.4 Pro73.3 Muse Wonder73.1 GPT-5.6 Sol73.0 GPT-5.5 Pro74.8

Lower price

Hy472.0 · $1.25/1M GPT-6.1 Terra77.1 · $5.00/1M Gemini 4 Carbon82.7 · $10.00/1M

More from Anthropic

Claude Sonnet 564.2 Claude Fable 5.176.2 Claude Mythos 5.1Unranked Claude Fable LatestUnranked Claude Fable 574.0