Claude Opus 5.5 may respond best when users stop treating it like an older model. Anthropic says some familiar prompt habits add time or cost without improving results, while clearer requests can let the model handle more of the work in one go.
Drop prompts that tell the model to think harder
Anthropic advises removing phrases such as “think hard” and “think carefully.” The company says Opus 5.5 reasons on its own, adjusting the depth of its reasoning to the task and the selected effort level. In Anthropic’s tests, removing those phrases made responses start faster, with no measurable drop in quality.
Those instructions came from earlier generations of reasoning models. At that time, Anthropic says, users could sometimes use them to coax better performance or influence how a request was routed. The company says that approach is no longer needed with today’s adaptive reasoning.
Begin at Medium effort and raise it when needed
Medium is Opus 5.5’s default effort level, and Anthropic recommends starting there. In the company’s internal tests, Opus 5.5 at Medium matched the performance of Opus 5 at High. Higher effort settings cost more and do not always lead to better answers.
If Medium does not meet the task’s needs, users can switch to High in the model menu next to the send button. Anthropic’s advice is to make that change based on the result, rather than automatically using a higher setting for every request.
The company also says teams that reserve Opus for only complex work may want to revisit that practice. A more capable model may complete a task in fewer steps, which can partly offset its higher per-token price. Anthropic says people who regularly raise Sonnet to Extra High or Max effort end up paying about as much as they would with Opus 5.5.
Give the full task in one request
Instead of sending instructions one at a time, Anthropic recommends outlining the desired result in a single message. Include the goal, the format you want, and when the model should stop to ask questions. This gives Opus 5.5 a clearer target from the beginning.
Anthropic says the model is better than its predecessor at carrying multi-part tasks through. It can take on some of the back-and-forth that users previously managed through follow-up messages, and may be able to guide itself from the initial request.
That approach also applies when you need a finished deliverable. If you want a completed document or spreadsheet, ask for the finished item directly instead of requesting a draft first. Anthropic says Opus 5.5 can often move from the request to the finished product in one pass.
Use files and precise design instructions
For diagrams and tables, Anthropic recommends attaching an image or file rather than manually typing the contents into a prompt. The company says older workarounds, such as transcribing chart values, are no longer necessary. It also reports improved image reading: even at the lowest effort level, Opus 5.5 read values from dense charts more accurately than Opus 5 at the highest effort level.
For design requests, specific direction is more useful than a broad ban. Anthropic suggests naming details to avoid, such as a cream-colored background or pill-shaped buttons, instead of simply asking for no “generic design.” General instructions, the company says, may just replace one default look with another.
Finally, users who connect Claude to their own files and systems through connectors can get more tailored results, according to Anthropic. Those connections let the model retrieve relevant information on its own. Taken together, the guidance points toward a simple shift: state the outcome clearly, provide useful source material, and adjust effort only when the task calls for it.