Claude Opus 5.5 removes disabled thinking and changes how agent turns end

Sources: primary, discussion
Summary: Thinking is always on and a disabled thinking type is no longer accepted, so integrations built for thinking-off should start at low effort and raise max_tokens, because thinking counts against the limit even when it is not returned. Default effort is medium rather than Opus 5's high, and Anthropic reports medium matching or beating Opus 5 at high on coding and knowledge work, so carrying the old value over produces longer turns and more output tokens. Progress notes now arrive as progress-update thinking blocks rather than text blocks and are empty at the default display setting, so a client that renders only text blocks looks silent through a long agentic turn, and a new reasoning_extraction refusal category declines prompts that push the model to reproduce its internal reasoning, with server-side fallback returning those declines rather than retrying them.
Why it matters: Long tasks can end a turn with text and an end_turn stop reason that an unattended loop reads as completion, and Anthropic's remedy is an explicit task checklist plus a bounded number of automatic continuations rather than unlimited retries.
Follow-up: Read the separate migration guide and what's-new page the prompting guide references, which list breaking changes not covered by the page verified here, and establish a release date.