Anthropic
Anthropic

Claude Sonnet 4

Release Date 2025-05-22(Knowledge cutoff 2025-01-31)1M IN · 64K OUTProprietary Model

Claude Sonnet 4 was Anthropic's balanced model for coding, reasoning and tool use, and is now retired on Anthropic-operated APIs. Anthropic described better coding and instruction following than the previous Sonnet, with a choice between responsive answers and extended reasoning for more demanding tasks. Tool use during reasoning and saving important facts to files required connected tools and storage rather than independent memory of all work inside the model. Current AIB comparisons place it on the weaker side for documents and automation, separate from its strong position when released. New integrations should follow current Sonnet successor and migration guidance, while users of external clouds should check their provider's schedule and test results after migration because availability can differ.

API cost

Official price
$3IN$15OUT
SourceAnthropic official docs

Output Speed

Slowper 2,000 tokens
45tok/s
First Output 0.8stotal: 45.5s
SourceArtificial AnalysisLast updated

User Comparisons

Response Speed
Total 6.9s
47 tok/s
SourceAIB Inc.Last updated

Expert Performance Evaluation

140284.2%18.9

More about Arena Text

16.7%64.6%1h 15m33.1%43.9%

More about τ³-Banking

70.3%138161.2%35.0%46.9%44.5%

By effortNone 44.0%Thinking 70.3%

More about Long Context Reasoning

46.6%Reasoning effort: Budget 2K10.3%89.7%58.9

More about Deep Research Bench

147565536.3%31.1%37.665.5%40.0%46.1%4.9%61.3%1446

More about Arena Coding

1399142154.7%139614168.1472.4%1389

More about Arena Writing, Literature & Language

10.7%0.3%140477.7%71.1%74.374.3%1416141299.1%84.4%4.1%0.0%40.0%5.9%45.5%3.1%77.1%29.0

By effortNone 4.3%Thinking 10.7%

More about Humanity's Last Exam

119134.0%37.0%44.8%

More about Arena Vision

Specialized fields

13381343143314341416140635.0%17.9%56.1%28.6%20.4%66.3%28.6%90.6%90.0%91.3%97.8%81.5%28.4%79.2%89.3%74.6%6.0%24.6%

More about Arena Korean

OddBench

30.0%Reasoning effort: High32

By effortNone 29.0%High 30.0%

More about BullshitBench v2

See each evaluation’s detail page for its source and last updateData corrections & removal requests