Anthropic
Anthropic

Claude Sonnet 4.5

Release Date 2025-09-29(Knowledge cutoff 2025-01-31)1M IN · 64K OUTProprietary ModelAPI available

Claude Sonnet 4.5 is an earlier Anthropic model for coding, computer tasks and multistep tool use, with retirement announced for its Anthropic-operated API. Anthropic reported improvements over the previous Sonnet in sustained work, reasoning and tool use, while the context-editing and memory tools introduced alongside it were API features for building longer workflows. Saving conversational records or rolling back work therefore requires product or development-environment support; calling the model does not automatically manage every record. Current AIB comparisons place it in the middle of the pack for documents and on the weaker side for coding and automation, separate from its leading results at launch. Long autonomous runs can accumulate mistaken assumptions or changes, so intermediate outputs and code need testing, and adoption should include a check of current availability and successors.

API cost

Official price
$3IN$15OUT
SourceAnthropic official docs

Output Speed

Slowper 2,000 tokens
46tok/s
First Output 0.9stotal: 44.4s
SourceOpenRouterLast updated

Expert Performance Evaluation

145587.5%20.7

More about Arena Text

24.5%$3,83978.1%43.8%2.1%2h 2m62.9%72.0%72.4%

More about τ³-Banking

72.3%146020.7%42.5%62.0%44.1%144954.5%55.3%

By effortNone 54.0%Thinking 72.3%

More about Long Context Reasoning

144930.7%52.6%43.9%33.6%16.0%115912.0%88.0%60.177.2%

More about Arena Text Factuality

0.0%151279655.8%35.6%52.1138571.4%45.7%71.3%47.7%14.7%149722.6%46.5%1.52×

More about Terminal-Bench 4.0

1454147757.3%1455146284.5%1444

More about Arena Writing, Literature & Language

17.8%1.1%142683.4%77.8%88.088.0%23.9%2.4%19.0%1471145097.7%15.2%4.2%63.7%13.6%54.3%6.0%10.0%17.0%83.2%38.8

By effortNone 7.2%Thinking 17.8%

More about Humanity's Last Exam

39.8%Reasoning effort: Budget 32K60.0%67.5%

More about VPCT

Specialized fields

1419141289.1%148314831473146460.0%28.9%2.4%90.8%66.3%61.8%31.1%29.0%69.8%31.1%91.0%67.5%73.5%76.2%70.9%94.0%85.8%93.1%80.8%015717786.0%81.0%70.0%42.0%13.0%84.0%89.4%96.0%76.0%72.0%75.0%66.0%1.6%75.8%32.1%79.5%89.0%75.1%62.9%29.2%84.0%5.0%9.5%

More about Arena Korean

OddBench

79.0%Reasoning effort: High37.3%33325566150

By effortNone 74.0%High 79.0%

More about BullshitBench v2

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