Anthropic
Anthropic

Claude Sonnet 4.6

Release Date 2026-02-17(Knowledge cutoff 2025-08-01)1M IN · 64K OUTProprietary ModelAPI available

Claude Sonnet 4.6 is Anthropic's model for coding, computer use and document work involving long inputs. Anthropic reported more consistent code, better instruction following and stronger planning and computer operation than the preceding Sonnet, making it an alternative to the contemporary Opus on some tasks. Reasoning settings can balance speed and analytical depth when reviewing codebases or refining documents, but accepting a long input does not ensure every detail is used correctly. Current AIB comparisons place it in the upper tier for documents, the middle of the pack for coding and automation, and on the weaker side for vision. Screen-based actions remain sensitive to interface changes and misread information, so users should test real work samples and check tool permissions and execution results.

API cost

Official price
$3IN$15OUT
SourceAnthropic official docs

Real-world usage

Outside weekly top 50
773B8/24
1.2T8/31
699B9/7
378B9/14
339B9/21

Output Speed

Very slowper 2,000 tokens
43tok/s
First Output 0.8stotal: 47.3s
SourceOpenRouterLast updated

User Comparisons

Writing Style
Answer length
Shorter side
bottom 32%
Vocabulary diversity
Average
bottom 50%
Sentence rhythm
Very steady
bottom 13%
Language mixing
High
top 36%
Reasoning share
Minimal
3% of total
Lies (Hallucination)
Honest answers
100%
Correct 0Honest don't-know 3Lies 0
Response Speed
instability ±30%
First response 3.4sTotal 14.8s
51 tok/s
16.7ms per token
User Reviews
  • •

    Claude is the best!

  • •

    A reliable AI that always delivers trustworthy results

  • •

    Easy-to-read and accurate answers. High stability.

  • •

    Even though it already knew the correct answer from its training data, it doubted its own accuracy regarding exact dates and ultimately answered "I don't know." It showed a trustworthy response.

  • •

    The model on the other side was chosen as such a low-quality one, so the response speed was way too different and the output had a lot of broken text.

SourceAIB Inc.Last updated

Expert Performance Evaluation

147273.030.1

More about Arena Text

34.4%$7,20479.5%61.3%41.5%13.3%72.1%

More about τ³-Banking

80.0%148143.0%36.6%51.0%65.3%25.8%38.5%148278.063.3%18.0%5.0%

By effortLow (no thinking) 69.3%High (no thinking) 68.3%Max 80.0%

More about Long Context Reasoning

146135.5%54.9%82.1%49.0%122010.6%89.4%59.691.1%

More about Arena Text Factuality

3.0%152842.6132758.1%71.2%53.0%63.0152250.1%75.2%24.3%29.9%66.1%151479.30.5%51.5%53.4%48.8%

More about Terminal-Bench 4.0

1455148056.6%1449147476.11197144563.2

More about Arena Writing, Literature & Language

33.6%3.1%146187.5%85.8%87.045.0%1500148284.832.4%8.3%86.5%60.4%23.0%16.0%89.9%46.5

By effortLow (no thinking) 11.2%High (no thinking) 13.3%Max 33.6%

More about Humanity's Last Exam

12836.7%77.4%73.0%74.5%

More about Arena Vision

Specialized fields

1434144089.3%15031506148841.7%148465.2%23.6%56.2%32.6%76.4%32.6%92.0%75.3%75.0%75.7%89.0%87.5%98.4%82.6%98133203-0.0291.0%89.0%76.7%68.0%29.0%77.0%89.5%87.0%75.0%77.0%81.0%67.0%70.5%82.3%88.6%79.6%81.5%30.9%12.0%

More about Arena Korean

OddBench

91.0%Reasoning effort: High80.9%26348296689215

By effortNone 89.0%High 91.0%

More about BullshitBench v2

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