Anthropic published prompting guidance for Claude Opus 5.5 on Sept. 22, and the headline instruction for developers is blunt: don't reuse the effort setting you ran on Opus 5. The new model defaults to medium effort, one level below the high default on its predecessor, and Anthropic's own testing shows Opus 5.5 at medium effort matches or surpasses Opus 5 at high effort on coding and knowledge work.

That shift matters for anyone with production apps. An application that never sets effort explicitly will run at medium on Opus 5.5 without any code change — but an app configured for Opus 5's high default will now spend more than it needs to on many tasks.

The guide, covered by Matt Southern at Search Engine Journal, positions effort as the first setting to adjust when balancing quality against speed and cost. For developers looking to cut thinking time, Anthropic recommends trying lower effort settings before making prompt changes, and reserving the higher levels — xhigh and max — for tasks where they genuinely move the needle on quality.

Thinking can no longer be turned off

There is one hard breaking change. Thinking cannot be disabled on Opus 5.5 at all; requests that try get an error. Opus 5 allowed exactly that at high effort or below.

This has a downstream consequence the guide flags directly: developers who used an output cap (max_tokens) with 'thinking off' on Opus 5 may see responses cut off on Opus 5.5, because hidden thinking consumes part of that cap even when it isn't visible in the output.

Despite the changes, Anthropic says Opus 5 prompts should still work well without edits, and its earlier Opus 5 guidance remains a reasonable starting point.

Anthropic suggests removing 'think carefully' from chat prompts

The guide advises chat application developers to remove system-prompt lines instructing Claude to think carefully before responding. The rationale: Opus 5.5 decides for itself how much to think, with effort serving as the main control.

Notably, Anthropic's own documentation includes that kind of line. For Claude Opus 4.7, when effort has to stay low to keep responses quick, the company's effort page suggests adding guidance such as "This task involves multistep reasoning. Think carefully before responding."

When Anthropic tested removal in a chat product, replies started sooner with "no clear decline in the quality of the reply," according to the guide.

Agent guidance: time budgets and tagged pasted text

For agent teams, the guide recommends setting a time budget based on how long tasks might take, with Opus 5.5 tracking elapsed time itself. Anthropic's tests found that small agent groups using time signals finish research tasks faster than a single agent working without signals, while achieving answer quality comparable to solo agents.

The time budget is a suggestion, not a hard constraint. The guide notes that setting a strict timeout can help, but the model may perform slightly less thoroughly under time pressure.

For apps handling pasted text from external sources like emails, Anthropic suggests wrapping it in tags with a random ID and adding a system note on how to manage tagged text. The tags encourage careful handling, but because they are plain text that can be copied, they provide only one layer of protection against prompt injection.

A pattern of migration guidance

The Opus 5.5 guide follows a broader pattern at Anthropic of asking developers to re-check configurations carried over from older models — workarounds for reading charts and screenshots, or instructions written for running with thinking off. Earlier this month, as Roger Montti covered, the Fable 5.1 guide asked developers to revisit formatting rules.

Frontend guidance in the new document is similarly specific. Anthropic recommends setting explicit styles to avoid defaults like cream backgrounds and pill-shaped buttons, observing that vague requests to avoid a generic AI look often just swap one default for another.

What to monitor

Developers migrating to Opus 5.5 should audit three things first: their effort settings, any max_tokens caps sized for thinking-off usage, and system prompts containing think-carefully instructions. Per the guide, teams running chat applications should A/B test with and without the think-carefully line, and agent teams should measure whether time budgets preserve answer quality in their own workflows.