Skills
Opus 5.5 effort migration: start at medium, set max_tokens to 128K, and use per-message effort changes to keep the cache
On Opus 5.5 thinking is always on, the default effort drops from high to medium, and turning thinking off or setting a thinking budget is rejected. Forced tool_choice any or a named tool now returns a 400 error. Anthropic says medium matches Opus 5 at high on coding. Keeping your old effort value produces longer turns, and thinking counts toward max_tokens, so size max_tokens up to 128,000. Changing top-level effort between requests invalidates the prompt cache, so use the per-message effort beta instead. Integrations that ran with thinking disabled should start at low effort and remove any show-your-reasoning instructions, which can now trigger a reasoning_extraction refusal.
↳ Follow the thread