Anthropic released Claude Opus 5.5 on September 22 and Sonnet 5.5 on September 28. Its launch announcements position Sonnet for well-defined everyday tasks and Opus for work requiring sustained judgment. The releases are available across Anthropic’s platforms and major cloud partners. Haiku 5.5 was still described as forthcoming in the checked announcements; it should not be reported as already released.

Sonnet’s launch page claims output generation more than 30% faster than Sonnet 5 and up to 30% lower cost per task in company testing. Its standard token prices did not fall: input remains US$2 and output US$10 per million tokens. Fewer tokens can reduce a task’s bill even when the unit price stays unchanged. Those results do not establish identical savings on a reader’s workload.

Opus 5.5’s announcement lists input at US$4 and output at US$20 per million tokens, each 20% below Opus 5. Anthropic also increased five-hour usage limits on several paid plans. API token prices and subscription usage allowances are different measures: neither is an unlimited-use promise. The company reports improvements in writing and complex work; Asal Sach has not independently benchmarked them.

Anthropic’s current model documentation lists a one-million-token context and 128,000-token maximum output for both models. They accept text and images and return text. API effort defaults differ: medium for Opus, high for Sonnet. An effort setting changes how much reasoning the model does, so a model-name-only comparison can hide meaningful differences in time and cost.

The Sonnet migration guide warns that older settings are not directly equivalent after upgrading. Compared with Sonnet 4.6, Sonnet 4.5 or Haiku 4.5, the same text can use about 30% more tokens with the newer tokenizer. Higher-resolution image handling can also increase image-token costs. Developers should measure real requests rather than assume a new version will always be cheaper.

For everyday bug fixes or a clearly specified document, Sonnet is a sensible candidate to test; for a difficult codebase change or an open-ended analysis, compare Opus as well. Anthropic’s model-selection guide recommends testing real prompts and edge cases and adjusting effort. Keep source checks, code tests and human review in the workflow. This is practical selection guidance, not evidence that either model is error-free. Official sources were checked on October 3.