WORLDTECHAnthropic urges Claude Opus 5.5 users to drop outdated prompt phrases
Anthropic says habits carried over from early reasoning models are holding back Claude Opus 5.5, and has published a guide telling users to abandon them. The company has also made Opus 5.5 the default model on all paid plans.
Cutting phrases and starting at Medium
According to the company, Opus 5.5 always reasons by itself and adjusts how deeply it reasons to the task and the chosen effort level. The guide's main advice is to strip out wording such as "think hard" or "think carefully" and to begin at the Medium effort level.
Anthropic dates these phrases to the early period of reasoning models, before adaptive reasoning became the norm. In its own tests, responses written without them began sooner, and the company reports no measurable difference in quality.
Medium setting compared with the predecessor
Anthropic says that in internal tests Opus 5.5 on Medium performed on a par with its predecessor Opus 5 on High. Higher settings cost more and do not always deliver better output, so the company advises moving to High only when Medium proves insufficient. Users can change the setting in the model menu beside the send button.
The model is now the default on Pro, Max, Ultra and Team plans. Anthropic contends that many users restrict it with instructions written for older models.
Implications for Sonnet and Opus users
Anthropic suggests that teams who have relied on the cheaper Sonnet and kept Opus for demanding projects should revisit that arrangement. A more capable model can complete a task in fewer steps, which may partly offset its higher per-token price. The company notes that anyone regularly running Sonnet at Extra High or Max effort ends up spending roughly what Opus 5.5 would cost.
Single-message goals and image input
The guide also recommends setting out the wanted result in one message rather than giving instructions piecemeal. That message should state the objective, the output format and the point at which the model should come back with questions. Anthropic says Opus 5.5 is better than its predecessor at carrying multi-part tasks through.
For diagrams and tables, the company proposes attaching them as images or files instead of retyping them. It also states that image reading has improved considerably in Opus 5.5.
Open questions
- Do the internal test results hold up in independent evaluations?