Claude Sonnet 5.5 released
On 28 September 2026 Anthropic released Claude Sonnet 5.5, the second model of its 5.5 family. It states that the model runs more than 30% faster than Sonnet 5 and costs up to 30% less per task, at an unchanged list price of $2 and $10 per million input and output tokens. It is the first Sonnet to launch with cyber safeguards and fallbacks, and with classifiers against reasoning extraction.
Why it matters
The mid-priced tier receives the same efficiency and safeguard changes as the top one, six days after Opus 5.5. The jump Anthropic reports on one agentic coding test, from 10.3% to 70.6%, is very large; the same table lists Opus 5.5 at 66.4%, and the page does not comment on that order.
Anthropic's own figures. Terminal-Bench 4.0: Sonnet 5.5 70.6%, Sonnet 5 10.3%, Opus 5.5 66.4% (at xhigh effort, its highest result). GDPval-AA v2.1: 1844 against Opus 5.5's 1846 and Sonnet 5's 1449; Artificial Analysis ran it on a pre-release deployment with a bug in structured outputs that, the page says, could degrade responses and has since been fixed. FrontierCode 1.1: 46.2% at Max effort, lower than at Xhigh; a footnote explains that at Max the model more often ran the code-review skill, which splits the work across subagents, and in two cases examined by Cognition this led to a timeout or to edits beyond the task. OSWorld 2.1 (partial score): 80.1% against 57.0% for Sonnet 5. The company says it is the first Sonnet to beat Pokémon Red working only from screenshots. Price and speed. $2 / $10 per million tokens and $0.20 for cache reads; output more than 30% faster; medium effort is the default in Claude Code and the apps. Available on the Claude API as claude-sonnet-5-5 and on the large clouds. The launch day, 28 September, comes from the API release notes entry. Safeguards. Sonnet 5.5's cyber capabilities are comparable to Opus 5's, so it is the first Sonnet to get cyber safeguards: higher-risk cyber tasks visibly fall back to Sonnet 5; the biology safeguards are as in Sonnet 5; the Cyber Verification Program is to be expanded. It is also the first Sonnet with classifiers against reasoning extraction and expanded 'preserved thinking'. Anthropic says Opus 5.5 remains clearly stronger at open-ended work requiring sustained judgement. What the record does not claim. All figures are measurements by Anthropic and its partner evaluators; no independent repetition was read. Haiku 5.5 is promised 'in the coming weeks'; its release is not part of this record.