Prompting Claude Opus 5.5: Anthropic's guide for developers

Anthropic has published a prompt-engineering guide for Claude Opus 5.5 in the Claude Platform documentation. The company says the model produces output tokens more than 30% faster than Claude Opus 5 and finishes the same task with fewer tokens, while prompts written for the older model should keep working unchanged.

Effort is the main control

Thinking is always on in Claude Opus 5.5, so effort is the first setting to adjust. The default is medium, where Claude Opus 5 defaulted to high, and Anthropic advises setting the level explicitly and testing several values against your own evaluations: in its testing, Opus 5.5 at medium matched or exceeded Opus 5 at high on coding and knowledge-work evaluations.

At a given level the model thinks more per turn than Opus 5, especially at xhigh and max, so developers should leave room in max_tokens for thinking tokens, which count toward the limit even when they are not returned; Anthropic used the model's maximum of 128,000 for long agentic coding turns. To get less thinking, lowering effort works more reliably than prompt instructions, and changing the top-level effort value between requests invalidates the prompt cache.

Agentic runs and harness changes

Opus 5.5 no longer accepts disabled thinking, which Opus 5 allowed at high effort or below. Integrations that ran with thinking off should start at low effort, remove instructions that asked the model to write its reasoning into the response, and read responses by block type.

In unattended agents, a turn that ends with text and no tool call is now a progress report rather than proof that the task is finished. Anthropic recommends keeping the task's parts in a checklist the model updates and sending a short continuation message that names the open items.

Between tool calls the model writes short progress updates, which arrive as progress-update thinking blocks with empty text by default, so clients must request the updates display setting.

Safety and visual inputs

Claude Opus 5.5 runs safety classifiers for biology, cybersecurity and reasoning extraction. A decline returns stop_reason "refusal" with a stop_details object naming the category, and such requests can be retried on a fallback model, except declines for reasoning extraction.

For dense charts, diagrams and screenshots the guide suggests higher-resolution images and image tools such as PIL and OpenCV in a container.