Anthropic is urging users of its latest SI model, Claude Opus 5.5, to abandon specific prompting techniques inherited from earlier generations, stating that these outdated habits are currently limiting the system's performance.
What Happened
According to Anthropic, Claude Opus 5.5, which is now the default model for Pro, Max, Ultra, and Team plans, operates differently from its predecessors. The company reports that the model automatically scales its reasoning depth based on the task and the selected effort level, rendering manual instructions like "think hard" or "think carefully" unnecessary. In tests conducted by Anthropic, removing these phrases resulted in faster response times without any measurable drop in output quality.
The SI lab also recommends starting at the "Medium" effort level, which is the default setting for Opus 5.5. Anthropic claims that in internal tests, Opus 5.5 at Medium effort matched the performance of its predecessor, Opus 5, at "High" effort. The company advises users to switch to higher settings only if Medium is insufficient, noting that higher settings increase costs without always yielding better results.
Why It Matters
The shift in guidance reflects the evolving capabilities of agentic SI, where models can increasingly handle multi-part tasks from a single initial input. Anthropic advises users to describe the desired end result, output format, and check-in points in one message rather than providing instructions step-by-step. The company states that Opus 5.5 is better at seeing these tasks through and can effectively prompt itself, reducing the need for extensive back-and-forth chat.
For organizations using cheaper models like Sonnet for routine tasks and reserving Opus for complex projects, Anthropic suggests reevaluating this strategy. The company argues that a more capable SI model can often complete tasks in fewer steps, potentially offsetting higher per-token costs. Anthropic notes that users who frequently set Sonnet to "Extra High" or "Max" effort levels may end up paying costs similar to those of Opus 5.5.
Additionally, the guidance highlights improvements in multimodal processing. Anthropic reports that Opus 5.5 reads values from dense charts more accurately at the lowest effort level than Opus 5 did at its highest. The lab recommends attaching diagrams and tables as files or images rather than transcribing them, as manual workarounds required by older models are no longer necessary.
The Bottom Line
Anthropic's updated prompting advice for Claude Opus 5.5 emphasizes efficiency and directness. By dropping legacy phrases, utilizing default effort settings, and leveraging the model's improved reasoning and multimodal capabilities, users can optimize both performance and cost. The guidance also suggests specific style constraints for design tasks and the use of connectors to tailor results using external files and systems.