Anthropic is advising Claude Opus 5.5 users to remove prompting habits developed for older reasoning models. The company says phrases such as “think hard” and “think carefully” no longer improve quality because Opus 5.5 adjusts its reasoning to the task and selected effort level. In its tests, removing those phrases produced faster starts without a measurable quality loss.
Opus 5.5 is now the default model on Pro, Max, Ultra and Team plans. Anthropic recommends beginning with the default Medium effort and moving to High only when results require it. The company says Medium matched the previous Opus 5 model at High effort in internal testing, while higher settings cost more and do not always improve output.
Users should also provide the goal, output format and conditions for asking questions in one message rather than revealing instructions step by step. For charts and diagrams, Anthropic recommends attaching the original image or file instead of manually transcribing values. Finished documents or spreadsheets can be requested directly.
These recommendations are company guidance, not independent benchmark results. They suggest a practical change in workflow: begin with a complete specification and moderate effort, then spend additional reasoning budget only after observing a concrete shortfall.