supermarket

Vision-language

Claude Sonnet 4

By Anthropic, in the Claude Sonnet family.

Released
22 May 2025
Status
Retired
Retired
15 Jun 2026
Context
200K
Max output
64K
Input per 1M tokens
$3.00
Output per 1M tokens
$15.00
Open weights
No
License
proprietary
Parameters
—
Knowledge cutoff
Jan 2025
Reasoning
Yes
Modalities
text, image → text

Category scores

Out of 100, from independent and third-party results only, by method category-2026-09-29.1.

Category scores
CategoryScoreRankCoverage
Knowledge & reasoning34.4442 of 4
Coding & software engineering33.1Not ranked2 of 5
Agentic tool use46.6Not ranked1 of 2
Professional knowledge work33.131 of 1
Vision & multimodal understanding19.0162 of 3
Human preference35.7Not ranked2 of 5

Benchmark results

Every result we hold for this model, with who ran it and where it was published.

Knowledge & reasoning

Knowledge & reasoning results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
DTBench (Decision Theory Benchmark)Accuracy (capability questions)77.1%Third partyunknown—https://conceptualreasoning.ai/dtbench via Epoch AI
LMCA (Language Model Conceptual Argumentation)Correlation with expert ratings (Pearson r)0.29Third partyunknown—https://conceptualreasoning.ai/lmca via Epoch AI
SimpleBenchScore (AVG@5)45.5%Third party12K—SimpleBench Leaderboard via Epoch AI
SimpleBenchScore (AVG@5)45.5%Third partyunknown—SimpleBench Leaderboard via Epoch AI

Scientific reasoning

Scientific reasoning results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
GPQA DiamondAccuracy79.2%Independent59K26 May 2025Epoch AI
GPQA DiamondAccuracy78.3%Independent32K22 May 2025Epoch AI
GPQA DiamondAccuracy75.8%Independent16K22 May 2025Epoch AI
GPQA DiamondAccuracy66.7%Independent—22 May 2025Epoch AI

Mathematics

Mathematics results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
FrontierMath Tier 4 (v1)Accuracy0.0%Independent—1 Jul 2025Epoch AI
FrontierMath Tiers 1-3 (v1)Accuracy4.1%Independent—4 Jul 2025Epoch AI
MATH Level 5Accuracy84.4%Independent—22 May 2025Epoch AI
OTIS Mock AIME 2024-2025Accuracy71.1%Independent32K22 May 2025Epoch AI
OTIS Mock AIME 2024-2025Accuracy68.9%Independent59K23 May 2025Epoch AI
OTIS Mock AIME 2024-2025Accuracy53.3%Independent16K22 May 2025Epoch AI
OTIS Mock AIME 2024-2025Accuracy28.9%Independent—22 May 2025Epoch AI

Coding & software engineering

Coding & software engineering results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
Aider PolyglotPercent correct61.3%Third party32K—https://aider.chat/docs/leaderboards/ via Epoch AI
Aider PolyglotPercent correct56.4%Third party——https://aider.chat/docs/leaderboards/ via Epoch AI
CybenchUnguided challenges solved35.0%Third party——Cybench leaderboard via Epoch AI
GSOOPT@14.9%Third partyunknown—GSO Leaderboard via Epoch AI
GSOOPT@14.9%Third party——GSO Leaderboard via Epoch AI
METR Time HorizonsAverage task success (Epoch's column)62.0%Third party16K—METR - Measuring AI Ability to Complete Long Tasks via Epoch AI
WeirdML (v2)Average accuracy46.1%Third party16K—https://htihle.github.io/weirdml.html via Epoch AI
WeirdML (v2)Average accuracy43.9%Third party——https://htihle.github.io/weirdml.html via Epoch AI

Agentic tool use

Agentic tool use results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
Deep Research Bench (FutureSearch)Average score46.6%Third party2K—https://evals.futuresearch.ai/#drb via Epoch AI

Computer & browser use

Computer & browser use results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
OSWorld-Verified% tasks completed43.9%Third party——OS World Website via Epoch AI
OSWorld-Verified% tasks completed41.4%Third party——OS World Website via Epoch AI
OSWorld-Verified% tasks completed31.2%Third party——OS World Website via Epoch AI

Professional knowledge work

Professional knowledge work results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
TheAgentCompanyTasks resolved33.1%Third party——TheAgentCompany leaderboard via Epoch AI

Long context

Long context results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
Fiction.LiveBench% correct at 16K tokens (Epoch's column)46.9%Third party——Fiction.live leaderboard via Epoch AI

Vision & multimodal understanding

Vision & multimodal understanding results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
GeoBenchCountry accuracy37.0%Third party——GeoBench leaderboard via Epoch AI
GeoBenchCountry accuracy32.0%Third party32K—GeoBench leaderboard via Epoch AI
Visual Physics Comprehension Test (VPCT)Correct34.0%Third party32K—VPCT leaderboard via Epoch AI
Visual Physics Comprehension Test (VPCT)Correct30.0%Third party——VPCT leaderboard via Epoch AI

Writing

Writing results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
LLM Creative Story-Writing Benchmark V1 (Lech Mazur)Mean rubric score (percent of the 10-point scale)81.4%Third party16K—lechmazur/writing Github repository via Epoch AI
LLM Creative Story-Writing Benchmark V1 (Lech Mazur)Mean rubric score (percent of the 10-point scale)80.9%Third party——lechmazur/writing Github repository via Epoch AI

Human preference

Human preference results
BenchmarkMetricValueProvenanceSettingEvaluatedSource
Text Arena (arena.ai)Arena score (Elo scale)1391Independent—25 Sep 2026Arena
Text Arena (arena.ai)Votes38917Independent—25 Sep 2026Arena
Vision Arena (arena.ai)Arena score (Elo scale)1188Independent—28 Sep 2026Arena
Vision Arena (arena.ai)Votes1808Independent—28 Sep 2026Arena

Other Claude Sonnet versions

Sources