Introducing Claude Sonnet 5.5

What it claims

Sonnet 5.5 (claude-sonnet-5-5) targets well-scoped everyday work: bug fixing, documents, slides, spreadsheets and UI polish. Opus 5.5 stays the pick for complex, open-ended work requiring sustained judgment, and Anthropic says Opus 5.5 "remains clearly stronger" there despite close benchmark scores. Haiku 5.5 is announced for "the coming weeks".

Evidence

The page gives these figures beyond the headline numbers in the summary:

• GDPval-AA (44 occupations, nine industries): two points below Opus 5.5 and about 400 points above Sonnet 5. Artificial Analysis ran it on a pre-release deployment with a structured-outputs bug. Anthropic expects any effect to be small and to understate the score.
• CursorBench: Sonnet 5.5's best score is within about two points of Opus 5.5.
• Terminal-Bench 4.0: Sonnet 5.5 scores 70.6% against 10.3% for Sonnet 5. The Opus 5.5 comparison uses Opus at Xhigh effort.
• Low and Medium effort: on several benchmarks, Sonnet 5.5 at Low or Medium beats Sonnet 5's best score at about a tenth of the cost per task.
• Higher effort: at the top settings it performs comparably to Opus 5.5 at a similar cost. The page says it complements Opus best at lower effort.
• Long-horizon work: it is the first Sonnet to beat Pokémon Red from screenshots only. It is close to Opus 5.5 on computer use and chart recognition, and outperforms GPT-6 Sol on long-horizon knowledge work.
• Tool calls: in head-to-head runs it batched tool calls more than Sonnet 5, which meant fewer steps and lower cost.
• Alignment audit (about 1,850 scenarios): it matches or beats Sonnet 5 on most measures. Opus 5.5 is slightly better overall. Sonnet 5.5 is the least likely of Anthropic's models to probe container limits.

How to use it

• Pricing: $2/M input tokens, $10/M output tokens, $0.20/M cache reads. It is available on the Claude Platform, AWS, Google Cloud and Azure, with zero data retention available.
• Default effort differs by surface: Medium in Claude Code and the apps, High on the Claude Platform. Benchmark costs and your API bills will not match Claude Code behaviour unless you set effort explicitly.
• Breaking change if you run with thinking off: switch to the new between_tools thinking setting before moving to Sonnet 5.5. It keeps up-front thinking off. Details are in the migration guide.
• Preserved thinking: thinking is now bound to the account that created it. If you move conversations between accounts, including switching accounts mid-session in Claude Code, read the linked docs article.

Limits

• Max effort can score worse than Xhigh. On FrontierCode, Max effort scored lower than Xhigh. At Max, the model more often invoked Claude Code's multi-subagent code-review skill. In two cases Cognition examined, this caused a timeout or edits beyond the task's scope, which FrontierCode penalises. If your pipeline penalises scope creep, Xhigh may be the better ceiling than Max.
• Cyber fallback is visible. Higher-risk cybersecurity requests fall back to Sonnet 5, and the user can see it. Routine finding and fixing of bugs is unaffected. Security teams should expect fallbacks until the expanded Cyber Verification Program opens; the page says only "soon".
• Biology false positives. The biology safeguards are unchanged from Sonnet 5. Some microbiology and virology requests may be flagged in error. The Life Sciences Verification Program offers broader access.
• New anti-distillation classifiers block reasoning extraction. This is new for the Sonnet line.
• Benchmark caveats. GPT-6 Sol comparison scores may predate OpenAI's image-understanding bug fix. Artificial Analysis expects no major impact.
• Alignment coverage. Anthropic states the alignment evaluations may miss tendencies they have not found.