Anthropic
Anthropic

Claude Opus 4.6

Release Date 2026-02-04(Knowledge cutoff 2025-09-01)1M IN · 128K OUTProprietary ModelAPI available

Claude Opus 4.6 is Anthropic's model for large codebases, long source material and sustained multistep work. Anthropic reported better planning, code review and debugging than the previous Opus, with improvements that also support research, analysis, documents and spreadsheets. Adaptive reasoning can be combined with tools, but agent collaboration and file creation also depend on product features and the execution environment rather than the model alone. Current AIB comparisons put it in the upper tier for coding and research and the middle of the pack for documents and automation, which differs from its position in launch-time tests. Long inputs and self-review do not eliminate omissions or calculation errors, so users should check cited evidence, figures and actual code behavior before relying on the result.

API cost

Official price
$5IN$25OUT
SourceAnthropic official docs

Output Speed

Very slowper 2,000 tokens
38tok/s
First Output 1.4stotal: 54.0s
SourceOpenRouterLast updated

Expert Performance Evaluation

149674.531.9

More about Arena Text

$8,01892.1%75.8%4.2%11h 59m57.7%

More about Vending-Bench 2

78.0%150046.3%67.0%49.1%150869.960.0%60.7%20.7%17.0%

By effortHigh (no thinking) 67.0%Max 78.0%

More about Long Context Reasoning

1493Reasoning effort: High47.0%55.3%86.6%53.0%125212.2%87.8%59.992.5%

More about Arena Text Factuality

154749.099753.4%48.5%48.1153751.9%78.7%26.6%78.0%41.2%153478.257.6%79.8%33.3%44.1%1.47×

More about Arena Coding

1489150953.1%1478149883.3121086.7%147663.3

More about Arena Writing, Literature & Language

39.9%150489.6%94.4%89.366.0%26.8%50.0%1518151188.740.7%22.9%94.0%69.2%67.6%7.6%21.0%25.0%91.2%55.8

By effortHigh (no thinking) 19.1%Max 39.9%

More about Humanity's Last Exam

131228.3%77.4%73.9%

More about Arena Vision

Specialized fields

1459147991.1%152615341511150893.0%66.6%12.7%96.7%54.6%36.8%76.7%36.8%96.2%70.2%70.2%97.6%85.0%50211264-0.0134.5%86.7%72.3%73.1%83.9%89.9%81.3%45.2%47.0%12.0%

More about Arena Korean

OddBench

87.0%Reasoning effort: High92.1%33570892324

By effortNone 83.0%High 87.0%

More about BullshitBench v2

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