SYLIN

Claude Sonnet

by Anthropic · Model · Language & code models

mixed

Near the top in intelligence, but costly and beaten on value by cheaper rivals

Released with large agentic coding and knowledge-work gains over previous Sonnet, close to Opus 5.5; default on free plan

Why we say so

Artificial Analysis ranks Sonnet 5.5 second overall, ahead of GPT-6 Astra, but also finds it slightly behind Opus 5.5 and the costliest per task, using the most output tokens. The AI Search reviewer says GPT 6.1 Sol or Astra give far better performance per cost. Arena puts it third on web development, and Google's own table has it ahead on computer control and terminal coding.

Versions

VersionReleasedRank
Claude Sonnet 5.5LATEST
mixed Very capable, but expensive and outdone by Opus 5.5 and cheaper GPT models
Oct 4
Claude Sonnet 4.6
unproven No evidence of its own in this period

Evidence

  1. Nov 11
    #3 on Arena webdev (1772)
  2. Oct 4
    Ranks number two on the Artificial Analysis intelligence index, ahead of GPT-6 Astra
  3. Oct 4
    Slightly less intelligent than Opus 5.5 and costliest per task of ranked models, using the most output tokens
  4. Oct 4
    Reviewer: GPT 6.1 Sol or GPT-6 Astra give much better performance per cost
  5. Aug 28
    Used as the fact-extraction model in Lemmalog's memory benchmarks

Cost to run

Alternatives

Gemini HydrogenAt xhigh scores 54 on AA Intelligence Index, one point behind Sonnet 5.5 and GPT-6.1 Sol; about 32% of Sonnet's cost, 2.5x Sol's
Claude OpusOpus 5.5 at xhigh outscores Sonnet 5.5 at max with 40% lower cost per task: a higher tier beats more effort
GPT TerraCheaper than Claude Sonnet 5.5 but scores lower on Terminal-Bench; roughly 10x cheaper per the video
Claude FableFable 5 (Claude Code, high effort) topped the speedrun: record 2,726, 81.7% of the gap to human record closed, 8.7 agent-days
GPT LunaGPT Luna (listed as GPT-5.6 Luna) scored 3,110 in 1.9 agent-days, roughly matching Sonnet 5 at 3,105