Claude Sonnet 5.5 Review

Sonnet 5.5 arrived on September 28, 2026, six days after Opus 5.5, as what Anthropic calls the faster, lower cost complement to it: the model for clearly scoped everyday work such as bug fixes, documents, slides and spreadsheets. The price did not move from Sonnet 5, $2 in and $10 out per million tokens, and cache reads were halved to 10 cents. Anthropic's own testing has it more than 30 percent faster than Sonnet 5 and up to 30 percent cheaper per task, because it spends far fewer tokens. Box reported 2.4 times the speed on 12 percent fewer tokens, and Balyasny's finance suite fell from about 497,000 tokens per answer to 121,000.

The benchmark table is the surprise. On Terminal-Bench 4.0 it scores 70.6 percent at maximum effort, above Opus 5.5's 66.4 and far above Sonnet 5's 10.3. On CursorBench 4.0 it reaches 55.5 percent against Opus 5.5's 57.8, on OSWorld 2.1 it scores 80.1 percent against 81.8, and on GDPval-AA knowledge work it sits at 1844 against 1846. Anthropic is careful to say Opus stays clearly ahead on open ended work that needs prolonged judgment, and Humanity's Last Exam with tools, 64.5 percent against 67.7, is where that gap shows. Its own notes also record that at maximum effort the model sometimes split code reviews across subagents and then timed out or strayed out of scope.

The costs are the family's. Higher risk cybersecurity requests fall back to Sonnet 5 unless your organization is in Anthropic's verification program, some microbiology and virology questions are flagged by mistake, and teams that ran with thinking off have to adopt a new setting before they switch. It shipped on the Claude apps, the API and every major cloud on day one, with zero data retention available. For most coding and office work it does what Opus does for half the price, which makes coding an S. Critical thinking stays an A, a step behind Opus 5.5 and Fable 5.1 on the hardest reasoning, and the whole is an A.