Anthropic: Optimizing Claude Opus 5.5 While Avoiding Old Reflexes

Anthropic has released a series of recommendations for using Claude Opus 5.5 more effectively. The publisher claims that certain inherited habits hinder response times, while the model handles complex tasks better, including image reading and one-pass production.

Request the final deliverable and leverage enhanced image reading

Anthropic recommends directly asking for the desired final document or table, rather than first requesting a draft. According to the company, Opus 5.5 often transitions from the initial request to the final result in a single operation. The publisher also notes that the model's ability to read images has significantly improved: even with minimal effort, it can read values from dense graphs with greater accuracy than Opus 5 at its maximum effort.

Adjust the effort level to control cost and quality

Anthropic advises starting with the "Medium" effort level, which is the default setting for Opus 5.5. Based on its internal tests, this mode matches the performance of Opus 5 in "High." Higher levels involve greater costs and do not always guarantee improved results; therefore, it is suggested to select the "High" option only when necessary. Anthropic also recommends that teams who used Sonnet for routine tasks reconsider this approach, as a model offering better performance can complete a task in fewer steps, sometimes offsetting the higher per-token price. Frequent use of Sonnet with "Very High" or "Max" settings results in an overall cost similar to that of Opus 5.5.

Get to the point rather than conversing step by step

Anthropic suggests formulating the goal, the expected output format, and when the model should ask questions in a single message, rather than providing sequential instructions. Opus 5.5 handles multi-part tasks better than its predecessor and can manage exchanges previously conducted in chat, sometimes more efficiently, by prompting from a single initial input. The publisher also advises directly attaching diagrams and tables as images or files, as workarounds like manually transcribing values are no longer necessary.

Abandon prompt tics and connect your data

Anthropic believes that certain prompt habits slow down Opus 5.5 and recommends eliminating phrases like "think carefully" or "consider well." According to the company, the model reasons autonomously and adjusts the depth of its reasoning based on the task and effort level, and its tests show that removing these phrases speeds up responses without measurable loss of quality. These habits date back to the early days of reasoning models, about a year and a half ago, before the adoption of adaptive reasoning. Anthropic also suggests connecting files and systems, precisely describing the expected outcome, sharing diagrams, explicitly naming undesirable design styles, and providing one's own house style. For design, the publisher recommends giving specific examples, such as a cream background or pill-shaped buttons, as general indications like "no generic design" simply result in swapping one default style for another. Those who connect Claude to their own files and systems via connectors benefit from significantly more personalized results, as the model can then retrieve the necessary information itself. Opus 5.5 is now the default model for Pro, Max, Ultra, and Team offerings.