Summary
Anthropic has released Claude Sonnet 5.5, reporting output speeds more than 30% higher than Sonnet 5 and up to 30% lower cost per task. The company also reports a sharp increase on its Terminal-Bench 4.0 coding evaluation.
Anthropic announced Claude Sonnet 5.5 on September 28, 2026, positioning it as a faster, lower-cost model for everyday tasks, coding and document creation. The company says it generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task in its testing, while keeping the same published per-token rates as its predecessor.
The new model sits alongside Claude Opus 5.5, which Anthropic describes as better suited to complex work requiring sustained judgment. Sonnet 5.5 is aimed at more clearly scoped tasks, with Anthropic highlighting bug fixes, polished documents, slides and spreadsheets. The model is available across Anthropic’s platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.
Coding and work benchmarks
Anthropic reports a 70.6% score for Sonnet 5.5 on Terminal-Bench 4.0, an evaluation of complex, multi-step tasks carried out in a command-line interface. Sonnet 5 scored 10.3% on the same benchmark, while Opus 5.5 scored 66.4% at its highest reported effort setting. These are benchmark results reported by Anthropic, rather than a general measure of how well the models perform on every coding task.
The company also reports a score of 1,844 on GDPval-AA, which evaluates work tasks across occupations. That is close to Opus 5.5’s 1,846 and higher than Sonnet 5’s 1,449. Anthropic says GDPval-AA covers 44 occupations and nine major industries. On another knowledge-work evaluation, AA-Briefcase, Sonnet 5.5 scored 1,811, compared with 1,822 for Opus 5.5 and 1,359 for Sonnet 5.
Benchmark scores depend on the evaluation and the model’s effort setting. Anthropic says Sonnet 5.5 at lower effort settings can exceed Sonnet 5’s best scores on some tests while costing less per task. It also cautions that benchmark results capture only one aspect of model capability, and says Opus remains stronger on complex, open-ended work requiring sustained judgment.
Pricing, speed and safeguards
Sonnet 5.5’s listed prices are $2 per million input tokens, $10 per million output tokens and $0.20 per million cache-read tokens—the same rates as Sonnet 5. Anthropic attributes the lower per-task cost to the new model typically using fewer tokens to complete the same work; its testing found savings of up to 30% per task. The actual balance of cost and quality can also depend on the effort setting, which controls how long the model works and checks its responses.
Anthropic says Sonnet 5.5 improved on or matched Sonnet 5 on most measures in an automated behavioral audit of roughly 1,850 scenarios. It is launching with cybersecurity safeguards and fallbacks similar to those used for Opus 5.5, because Anthropic says its cyber capabilities are comparable to Opus 5’s. The company says routine software development remains available, while higher-risk cybersecurity tasks can fall back to Sonnet 5. Its biology safeguards are unchanged from Sonnet 5.
The release gives developers a faster Sonnet option at the same token prices, with the potential for lower costs when tasks require fewer tokens. Anthropic’s published comparisons indicate particularly large gains on its coding evaluation, while the company’s own assessment continues to place Opus ahead for open-ended work that calls for sustained judgment.