Instructions that are no longer needed
Anthropic has made Claude Opus 5.5 the default model on all paid plans, citing Pro, Max, Ultra, and Team, and has released a guide to correct prompting habits inherited from previous generations. According to the company, many users limit the model’s performance with outdated instructions. At the top of the list is the recommendation to eliminate phrases such as “think deeply” or “reason carefully”: Opus 5.5, Anthropic maintains, always reasons on its own and calibrates the depth of reasoning based on the task and the selected effort level. In the tests cited by the company, responses without these phrases start more quickly without measurable declines in quality.
These phrases date, according to Anthropic, to the early days of reasoning models, about a year and a half ago, when they could serve to steer the model’s behavior. What is new is that today that latitude is no longer seen as necessary: adjustment would take place through the effort level, not through exhortations in the prompt text.
Effort and costs: the internal comparison
The guide recommends starting from the “Medium” level, the default for Opus 5.5. Anthropic states that in internal tests Opus 5.5 at “Medium” matched the performance of its predecessor Opus 5 at “High.” Higher levels cost more and do not always produce better results, according to the company, which suggests switching to “High” only when “Medium” is not enough; the setting is changed from the model menu next to the send button.
On the cost front, Anthropic urges users to reconsider the combined use of the cheaper Sonnet for ordinary tasks and Opus for complex projects. The argument is that a more capable model completes a task in fewer steps, partly offsetting the higher price per token; users who often take Sonnet to the “Extra High” or “Max” levels would, according to the company, end up spending about as much as with Opus 5.5. These are the company’s statements, not data that can be independently verified from the available material.
Single prompt, attachments and direct requests
Anthropic recommends describing the desired result in a single message instead of proceeding through successive instructions, specifying the goal, output format and when the model should come back with questions. According to the company, Opus 5.5 follows multi-part tasks better than its predecessor and can autonomously handle the back-and-forth that previously took place in chat, in some cases going so far as to formulate its own prompts from a single initial input.
For diagrams and tables, the guide suggests attaching them as images or files instead of transcribing them: the workarounds required by previous models, such as manually transcribing the values of a chart, would no longer be necessary. Anyone who needs a finished document or spreadsheet should ask for it directly rather than a draft, because the model often reaches the complete product in a single pass. Anthropic also reports an improvement in image reading: even at the lowest effort level, Opus 5.5 reads values from dense charts with greater accuracy than Opus 5 at the maximum level.
Design and connectors
For design tasks, the company recommends explicitly naming unwanted styles, such as a cream-colored background or pill-shaped buttons, instead of generically banning “generic design”: vague indications, according to Anthropic, would merely replace one default style with another. For users who connect Claude to their files and systems through connectors, the company states that results are much more personalized because the model retrieves the necessary information on its own.
The performance and cost comparisons cited by Anthropic remain to be independently verified, as the company does not publish the test details in the material. The guide also describes a change in usage practices more than in the model’s capabilities: the company’s thesis is that more concise prompts and calibrated effort levels make better use of Opus 5.5, but the real extent of this effect for different workflows is not quantified in the source.
