Anthropic released Claude Sonnet 5.5 on September 28, 2026, calling it the second model in the Claude 5.5 family. According to Anthropic’s announcement, the model generates output 30%+ faster than Claude Sonnet 5 and is “the first Sonnet model to launch with cyber safeguards and fallbacks like those we’ve developed for our most capable models.”
On this page
Availability and model ID
Developers call the model as claude-sonnet-5-5. Anthropic lists it on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure, and in the Claude apps and Claude Code.
Cyber safeguards with a fallback
The safeguards change how the model handles security work. Anthropic says: “Users can still find and fix bugs in their code as part of routine software development, but higher-risk cybersecurity tasks will visibly fall back to Sonnet 5.”
In practice, ordinary debugging stays on Sonnet 5.5, while requests that Anthropic classes as higher-risk cybersecurity work are answered by Claude Sonnet 5, and the switch is visible to the user. Until now, that tier of safeguards had been reserved for Anthropic’s most capable models.
Preserved thinking expands
Anthropic says Sonnet 5.5 “expands preserved thinking, so Claude’s thinking cannot be decoupled from the account that created it.” The company notes that users who move conversations between accounts will see changes in behavior and points developers to a migration guide. In its Claude Opus 5.5 announcement, Anthropic describes preserved thinking as “the anti-distillation safeguard we introduced with Fable 5.1.”
Benchmark results
All scores below were published by Anthropic on the Sonnet 5.5 page. Settings are shown where Anthropic gives them.
| Benchmark | Claude Sonnet 5.5 | Setting |
|---|---|---|
| Terminal-Bench 4.0 | 70.6% | Default |
| FrontierCode 1.1 (Main) | 52.1% | Xhigh effort |
| CursorBench 4.0 | 55.5% | Default |
| GDPval-AA v2.1 | 1844 | Default |
| AA-Briefcase v1.1 | 1811 | Default |
| OSWorld 2.1 | 80.1% | Listed as “partial” |
On FrontierCode 1.1, Anthropic reports a lower score at max effort (46.2%) than at xhigh effort, which is the model’s default for that benchmark. Anthropic’s footnotes say FrontierCode penalizes out-of-scope changes to the code. Lab-published scores for Sonnet 5.5 and other models are tracked on our benchmark leaderboard.
What it changes
Sonnet 5.5 brings two controls that earlier sat only on Anthropic’s top models down to its mid-tier line: the cyber fallback and the account-bound thinking. For developers building security tooling on Sonnet, the fallback means some requests will be served by Sonnet 5 rather than Sonnet 5.5. For applications that move conversation state between accounts, the preserved thinking change is the one to test before switching model IDs.
Sonnet 5.5 is listed on our AI model release timeline, and service health for Claude is tracked on the Claude status page.
