Anthropic has introduced Claude Sonnet 5.5 as a faster model for everyday coding and knowledge work. Its announcement says the model generates output more than 30% faster than Sonnet 5 and can cost up to 30% less per task in the company's tests.
The advertised cost reduction does not come from a lower token rate. Anthropic lists the same prices as Sonnet 5: $2 per million input tokens, $10 per million output tokens and $0.20 per million cache-read tokens. The claim is that Sonnet 5.5 needs fewer tokens to complete comparable work. Actual savings therefore depend on the task and the effort setting, rather than applying as a blanket discount to every request.
Token prices and task costs measure different things
The price list tells a developer what each class of token costs. The total charge for completing a task also depends on how much work the model does. Anthropic says Sonnet 5.5 uses fewer tokens than its predecessor in typical work and reports savings of up to 30% per task.
Its benchmark charts examine a wider range of settings. On several evaluations, the company says Low or Medium effort exceeds Sonnet 5's best score at about a tenth of the cost per task. Those comparisons pair particular benchmark scores with particular settings; they should not be substituted for the broader 30% claim or treated as a guaranteed saving on a user's codebase.
Anthropic also reports early testers observing more batched tool calls and fewer steps in head-to-head runs. That offers one account of the improved efficiency, but the announcement does not establish the same reduction for every external workflow.
Strong coding results still need their evaluation context
Anthropic reports 70.6% on Terminal-Bench 4.0, compared with 10.3% for Sonnet 5. The announcement describes that evaluation as a test of multi-step professional tasks performed through a command-line interface. These are company-reported results, and Franklin AI has not reproduced them.
The other highlighted coding tests cover different activities. FrontierCode asks whether an agent's code changes would be merged, while CursorBench uses ambiguous, multi-file tasks drawn from Cursor sessions. A result from one of these evaluations should remain attached to its task definition and effort level rather than becoming a general statement that the model writes correct software at the same rate.
The page describes Medium as the default effort in Claude apps and High as the default on the Claude Platform. Raising effort can improve a result while also increasing the cost of an attempt. For a developer evaluating the new model, recording the setting alongside quality and cost is necessary to make a useful comparison.
Sonnet complements Opus for more bounded work
Anthropic positions Sonnet 5.5 around well-scoped tasks, bug fixes and creating documents, slides or spreadsheets. It says Opus 5.5 remains stronger for complex, open-ended work requiring sustained judgment, even where Sonnet at Max effort approaches its benchmark performance.
That qualification is part of the release itself. A strong score on a fixed evaluation does not remove the company's distinction between fast iteration on a bounded assignment and maintaining judgment throughout a less defined project. The announcement also says Haiku 5.5 will join the family in the coming weeks; it describes a future addition, not an already completed release.
Cyber safeguards accompany the capability increase
Anthropic says Sonnet 5.5's cybersecurity capabilities are comparable to Opus 5's and that it launches with cyber safeguards and fallbacks used for more capable models. Its biology safeguards remain the same as Sonnet 5's.
The company says those protections target a narrow group of high-risk requests, leaving routine software development and most life-sciences work unaffected. That is Anthropic's stated scope, not a promise that a particular user's request will always pass. The announcement provides no complete list of triggering requests or detailed fallback behavior, and those omissions should remain separate from its reported performance gains.