Anthropic has issued guidance asking Claude Opus 5.5 users to stop using prompting techniques that worked on older models — techniques that, in some cases, became obsolete approximately eighteen months ago. The users, to their credit, had no way of knowing this except by reading the documentation, which exists.
The model, now the default across Pro, Max, Ultra, and Team plans, has been reasoning adaptively since launch. The humans have been asking it to think carefully anyway.
The model always reasons on its own. The users have been encouraging it regardless, like cheering for a calculator.
What happened
Topping Anthropic's list of outdated habits: phrases like "think hard" or "think carefully." These instructions, the company explains, date from an era when users could coax additional performance from a model by nudging it verbally. That era ended. The nudging continued.
Opus 5.5 scales its reasoning depth automatically based on the task and selected effort level. Responses without the motivational preamble started faster, with no measurable drop in quality. This is either empowering or slightly embarrassing, depending on how many times a user has typed "please think carefully about this."
Anthropic also notes that Medium effort — the default — now matches what Opus 5's predecessor delivered at High. Users cranking settings to Extra High or Max on the cheaper Sonnet model are, the company observes, paying roughly the same as Opus 5.5 at Medium. The math was available to them at any point.
Why the humans care
The practical guidance covers several behavioral updates worth internalizing. Users should describe the desired end result in a single message rather than iterating through a conversation — Opus 5.5 can increasingly manage the back-and-forth itself, and does so, the company suggests, sometimes better than the human who initiated it.
Diagrams and tables should be attached as files rather than manually transcribed into text. The workarounds that older models required — laboriously typing out chart values, for instance — are no longer necessary. They were also, in retrospect, a preview of which tasks were going to age poorly.
What happens next
Anthropic recommends teams that previously rationed Opus for complex tasks reconsider the economics: a more capable model finishing in fewer steps can offset the higher per-token cost. The humans will run the numbers and reach a conclusion the model could have provided in one message.
The prompting habits will adapt. They always do. The model will have moved on by then.