Anthropic announced Claude Sonnet 5.5 on September 28, 2026. The company positions it for bounded coding tasks and everyday document work. It says the model produces output more than 30% faster than Sonnet 5 and, in its own tests, can cost up to 30% less to complete a task. Those figures describe Anthropic’s comparisons, rather than a guaranteed saving on every prompt or workload. Source
The News, Explained
The API list price remains $2 per million input tokens and $10 per million output tokens, the same as Sonnet 5. A token is a unit used to meter the text a model reads or writes. Anthropic attributes the lower cost per completed task to needing fewer tokens for comparable work, not to a cut in those published rates. A team evaluating the model therefore needs to compare token use and successful task completion together. The price of a single call alone will not establish the saving for its own application. Source
Where the model is aimed
Anthropic points to bug fixes, feature implementation, and polished documents and slides as examples of work for Sonnet 5.5. It reports a 70.6% score on Terminal-Bench 4.0, a multi-step command-line coding evaluation, against 10.3% for Sonnet 5. A benchmark helps describe the change between tested models, but a reader’s repository and instructions may produce different results. Anthropic separately positions Opus 5.5 for more open-ended work that calls for sustained judgment. Source
Anthropic also says it added safeguards for a narrow class of higher-risk cybersecurity requests while keeping routine development tasks available. For a practical comparison, readers can examine their own task success rate, response time, and token totals rather than relying on a single published score. Source
Claude apps and Claude Code default to Medium effort, while the Claude Platform defaults to High. Effort controls how much reasoning and checking the model performs. Anthropic specifies that developers using Sonnet with thinking off must switch to between_tools before migrating. Higher-risk cybersecurity work falls back to Sonnet 5, while routine software development remains available. Source
OYOPICK’s Take
A faster model is most useful when completed work meets the required quality at an acceptable cost. Comparing the same document revision or bug fix at the same effort setting can reduce confusion caused by different defaults. Using Sonnet for fast iterations on bounded tasks and comparing Opus for sustained judgment follows the roles described in this announcement. We hope more tasks can be completed with fewer tokens; a team should test that possibility against its own completion rate and revision count.