Anthropic Releases Claude Sonnet 5.5

Anthropic released Claude Sonnet 5.5 on September 28. The company reports over 30% faster output and up to 30% lower cost per task compared with Sonnet 5.

The savings come from using fewer tokens, rather than a price cut. API rates stay at $2 per million input tokens and $10 per million output tokens. Cache reads cost $0.20 per million tokens.

Anthropic reports a Terminal-Bench 4.0 score of 70.6%, compared with Sonnet 5's 10.3%. The company still recommends Opus 5.5 for complex, open-ended work requiring sustained judgment. Sonnet targets more clearly defined tasks.

GitHub confirmed general availability in Copilot. Its early testing found Sonnet 5.5 matched Sonnet 5 on coding tasks while taking fewer steps and using fewer tokens and tool calls. GitHub's findings describe efficiency in its own tests, not a universal improvement across every coding workload.

Sonnet 5.5 also adds cyber safeguards similar to those used for Anthropic's more capable models. Anthropic says the restrictions target high-risk requests and leave routine software development unaffected.

API migration involves more than changing the model name. Google Cloud's documentation confirms Sonnet 5.5 rejects the old disabled-thinking setting. The replacement, between_tools, skips extended thinking but returns progress updates between tool calls as thinking blocks. Applications must handle those blocks instead of assuming every response contains only text.

LiteLLM's release guidance identifies another compatibility change: forced tool selection returns an HTTP 400 error. Integrations using required or named tool choices need to switch to automatic selection and describe the intended tool use in the prompt.