Claude Sonnet 5.5 launched on September 28, 2026, with unchanged standard token prices and reported improvements in speed and task efficiency. Anthropic’s claim of up to 30% lower task costs refers to using fewer tokens to finish work, not a 30% reduction in the listed token rate.
Read the cost claim at the task level
Anthropic lists input at $2 and output at $10 per million tokens, with cache reads at $0.20 per million. It reports output generation more than 30% faster than Sonnet 5 and positions the model for well-scoped everyday work. Opus 5.5 remains its recommendation for more complex, open-ended judgment.
For a production workflow, the meaningful denominator is an accepted result. A cheaper individual response can become expensive if someone must repeat the task or repair its output. Conversely, a model with the same rate can reduce costs if it needs fewer attempts. Include tool calls, retries, review labor, and failed runs when comparing the two versions.
Use a migration test that resembles real coding
Start with a representative group of bug fixes and bounded changes from your repository. Keep the prompts, available tools, and acceptance checks stable. Review changes without revealing which model produced them where practical, then compare passing results, elapsed time, and total spend.
A benchmark score does not reveal whether an agent preserves your authorization rules or understands an unusual build system. Include at least one task where the correct answer is to avoid a change, one that needs clarification, and one that exercises a dependency boundary. These cases test engineering judgment that a straightforward implementation demo can hide.
Keep effort settings and safety behavior in the comparison
The announcement describes different default effort settings in Claude applications and the developer platform. A comparison that changes effort, tools, and model at once cannot tell you which change helped. Use the same setting first, then investigate whether a lower-effort configuration meets the same acceptance bar.
Anthropic also identifies cyber safeguards for this Sonnet release. Its Cyber Verification Program announcement explains vetted access for qualifying security work. Treat a legitimate security task blocked during evaluation as a configuration and eligibility question, rather than an invitation to remove controls. Ordinary coding and specialized security investigation should have separate evaluation cases.
When an upgrade earns its place
A sensible rollout moves one class of routine work first and preserves the previous configuration for comparison. Expand only when the accepted-result rate and operational measurements hold up. Sonnet 5.5 is a candidate for a better daily-work default; the vendor’s release does not establish that every existing workflow should change immediately.