Anthropic has published a prompting guide for teams moving applications from Claude Opus 5 to Opus 5.5. Its first recommendation is to set the model’s effort level explicitly and test it on the application’s own tasks. Opus 5.5 defaults to medium effort, while Opus 5 defaults to high, and Anthropic says the same level name does not imply the same amount of thinking across models.
That matters for both quality and cost. Anthropic says Opus 5.5 at medium matched or exceeded Opus 5 at high on its coding and knowledge-work evaluations. It recommends reserving xhigh and max for tasks where a measured quality gain pays for the extra tokens and time. For an integration that previously disabled thinking, Anthropic suggests starting at low effort and measuring the result, since Opus 5.5 always has thinking enabled.
Check the request and the agent loop
The guide says thinking tokens count toward max_tokens, even when the application does not display them. A limit sized for Opus 5 with thinking disabled can therefore cut off an Opus 5.5 reply. Anthropic also says changing the top-level effort setting between requests invalidates the prompt cache. Its per-message effort option, currently in beta, can change individual turns while retaining that cache.
There is a separate failure mode for unattended agents. Opus 5.5 can finish a turn with a progress report and stop_reason: "end_turn" while work remains. Anthropic advises keeping an explicit task checklist and continuing the loop when a report arrives with unfinished items. It recommends a limit of two or three automatic continuations on the same stalled task.
Anthropic’s guide also says progress notes between tool calls arrive in thinking blocks. A client that renders only text blocks may appear silent. The display: "updates" setting exposes summaries of those notes.
