Anthropic launches Claude Sonnet 5.5 with lower cost and faster output

Anthropic has released Claude Sonnet 5.5, the second model in its Claude 5.5 family. The company announced it on September 28, 2026, describing it as a faster, cheaper complement to Claude Opus 5.5. Sonnet 5.5 is aimed at well-scoped everyday work, software fixes, documents, slides and spreadsheets, while Anthropic says Opus remains the stronger option for complex, open-ended work that requires sustained judgment. The model is available through Anthropic’s platform and across Amazon Web Services, Google Cloud and Microsoft Azure.

Anthropic reports that Sonnet 5.5 produces output more than 30 percent faster than Sonnet 5 and costs up to 30 percent less per task in its testing. Its list price is unchanged at $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20. The company says the lower task cost comes from needing fewer tokens for the same work, rather than from a lower token price alone. The API model identifier is claude-sonnet-5-5.

The release includes large provider-reported benchmark gains. On Terminal-Bench 4.0, Anthropic reports 70.6 percent for Sonnet 5.5 against 10.3 percent for Sonnet 5. On CursorBench 4.0, it reports 55.5 percent against 34.1 percent. Its GDPval-AA score is 1844, compared with 1449 for Sonnet 5, while Opus 5.5 remains slightly ahead on several tests. These figures come from Anthropic’s own evaluation setup and should not be read as a universal ranking. CodeRabbit independently tested the model on 44 real pull requests and reported more catches in roughly half the time, offering early practical context rather than a complete benchmark.

Safety is part of the launch. Anthropic says Sonnet 5.5 has cybersecurity safeguards and fallbacks because its cyber capabilities are comparable to Opus 5, plus biology safeguards and classifiers intended to limit industrial-scale reasoning extraction. The company found no evidence in its audit that the model pursues goals conflicting with user intent, but also notes that evaluations cannot catch every failure.

For users and developers, the significance is practical: a model positioned between ordinary assistants and the most expensive frontier systems is becoming faster, cheaper and broadly deployable. That could make coding agents and document workflows more economical, while the model’s limits, safeguards and provider claims still need testing against each team’s real tasks.