Anthropic is telling developers to control Claude Opus 5.5's thinking with the model's effort setting instead of filling prompts with instructions about how hard or how long it should reason.
In its official Opus 5.5 prompting guide, Anthropic says the model always has thinking available and decides how much to use. For applications where latency matters, the company recommends lowering effort first, starting at low and measuring the quality tradeoff.
The popular advice is broader than Anthropic's test
The guidance spread on social media as a call to delete familiar phrases such as “take a deep breath” and “think step by step.” Anthropic's documented finding is more specific. It says that, in one chat product, removing a system-prompt line telling the model to think carefully made replies start sooner with “no clear decline” in reply quality.
That result does not prove that every reasoning cue hurts every task. It does support testing whether old prompt scaffolding still earns its latency and token cost now that effort is a separate API control.
Anthropic also recommends removing instructions that ask the model to write out internal reasoning as a substitute for built-in thinking. For integrations that previously disabled thinking, the company suggests using summarized thinking blocks when developers need visibility and testing lower effort levels rather than demanding a step-by-step transcript in the response.
The practical takeaway is to audit prompts written for older model behavior, keep rules that define the task or output, and test whether instructions about the model's internal process still improve results. Anthropic repeatedly recommends measuring changes against an application's own evaluations instead of treating a single prompt formula as universal.
