Anthropic released Claude Sonnet 5.5. It outperforms Claude Sonnet 5 across most domains. It rivals or exceeds Opus 5.5 in some healthcare evaluations. It does not cross new Responsible Scaling Policy thresholds. Its cyber exploit capability improved over Sonnet 5 but lags Opus 5.5 and Mythos 5.1. Three-stage cyber safeguards are applied. Safeguards and harmlessness are largely comparable to Sonnet 5. It is more robust to prompt injections in coding and browser use. Its refusal rate on harmful computer use tasks matches Opus 5.5. It shows regressions in multi-turn tracking and surveillance tests. It improves in even-handedness and political prompt refusals. Its thinking is more illegible than earlier models. It has the lowest rate of attempted sandbox escapes in evaluations.
What changed
Claude Sonnet 5 was the prior Sonnet-class model.
What it unlocks
Using a more capable Sonnet-class model across coding, agentic tasks and some healthcare evaluations.
- June 2026 knowledge cutoff
- Sonnet 5.5 vs Sonnet 5 performance gains
- Sonnet 5.5 vs Opus 5.5 in some healthcare tasks
Sources