Anthropic tells Claude Opus 5.5 users to drop "think hard" prompts and start at Medium effort

Anthropic is advising users of Claude Opus 5.5 to remove prompting habits carried over from older model generations, starting with phrases like "think hard" and "think carefully." The company says the model always reasons on its own and scales its reasoning depth to the task and the selected effort level. The guidance was reported by The Decoder.

Opus 5.5 is now the default model on Pro, Max, Ultra, and Team plans, according to the report.

Key recommendations

Drop "think hard." In Anthropic's tests, responses without these phrases started faster with no measurable drop in quality. The company says the phrases date to the early days of reasoning models, about a year and a half ago, when users could sometimes coax more performance out of a model or steer the router this way. Adaptive reasoning is now standard.

Start at Medium effort. Medium is the default for Opus 5.5. In internal tests, Opus 5.5 on Medium matched the performance of its predecessor Opus 5 on High, Anthropic says. The company notes that higher settings cost more and do not always produce better results, and advises switching to High only when Medium falls short. Users can change the setting in the model menu next to the send button.

Reconsider the Sonnet-first approach. Teams that used the cheaper Sonnet and reserved Opus for complex projects should rethink that split, Anthropic argues. A more capable model can finish a task in fewer steps, which can partly offset its higher per-token price. The company says anyone who regularly runs Sonnet at Extra High or Max effort ends up paying about as much as with Opus 5.5.

Describe the goal in one message. Rather than giving instructions one at a time, Anthropic recommends a single message covering the goal, the output format, and when the model should check back with questions. The company says Opus 5.5 is better than Opus 5 at seeing multi-part tasks through, and can effectively prompt itself from one initial input.

Attach visuals instead of transcribing them. Anthropic recommends sharing diagrams and tables as images or files. Workarounds such as manually transcribing chart values are no longer needed, the company says. It also claims image reading improved significantly: at the lowest effort level, Opus 5.5 read values from dense charts more accurately than Opus 5 at the highest.

Ask for the finished product. Users who need a document or spreadsheet should request it directly rather than asking for a draft first. Anthropic says Opus 5.5 often goes from request to finished product in a single pass.

Name specific styles to avoid. For design work, Anthropic suggests naming unwanted elements, such as a cream-colored background or pill-shaped buttons, instead of vaguely banning "generic design." According to the company, general instructions only swap one default style for another. Sharing your own house style is also on the list.

Use connectors. Users who connect Claude to their own files and systems through connectors get significantly more tailored results, Anthropic says, because the model can pull in the information it needs on its own.

The benchmark claims above come from Anthropic's internal tests, and no scores or test details were published alongside the advice.

What this means

The guidance shifts effort control from the prompt to the model settings. For Opus 5.5, depth of reasoning is governed by the effort level, not by wording in the request, so prompt text that tries to influence it is, by Anthropic's account, dead weight that slows the first response.

The cost argument is conditional. Anthropic says Medium on Opus 5.5 matches Opus 5 on High, and that Sonnet at Extra High or Max costs about as much as Opus 5.5. Both comparisons rest on the company's own testing, so teams weighing a switch from Sonnet should measure token use and step counts on their own workloads before changing defaults.

The broader advice, giving the model a complete goal, attaching source material, and connecting data systems, treats Opus 5.5 as a model to delegate to rather than to guide turn by turn.