To use Claude Opus 5.5 well, start at its default medium effort and tune effort before anything else, leave room in max_tokens for thinking, and update a few prompt habits carried over from older models. That is the core of Anthropic's official "Prompting Claude Opus 5.5" guide in the Claude Platform docs, which also notes that existing Claude Opus 5 prompts should keep working without changes.
From Anthropic's testing: Opus 5.5 generates output more than 30% faster than Claude Opus 5, and at its default medium effort it matched or beat Opus 5 at high effort on multistep coding work, in fewer steps and with fewer tokens.
What is better in Claude Opus 5.5
- Knowledge work: much less likely to state an incorrect figure or cite the wrong source.
- Charts and screenshots: even at its lowest effort it read values off dense charts more accurately than Opus 5 at its highest.
- Long agentic work: it sustains multihour autonomous tasks better, and its progress updates say plainly what it did and what it needs.
7 tips from Anthropic's guide
- Treat effort as the main control. Thinking is always on, so start at medium (the default), set it explicitly, and test a few levels on your own tasks rather than copying your Opus 5 setting.
- Set max_tokens high enough. Thinking counts toward the limit even when you don't see it; for long agentic coding turns Anthropic found 128,000 worked well.
- Lower effort to get less thinking. It cuts cost and latency more reliably than prompt instructions do.
- In unattended agents, treat a text-only stop as a progress report, not proof the task is done. Keep a checklist and send a short nudge naming open items, but stop after two or three automatic continuations.
- For agents that work across email, documents and spreadsheets, add one sentence telling the model to explore the relevant sources before acting.
- In chat apps, remove old "think carefully before answering" lines. In Anthropic's testing, replies started sooner with no clear loss of quality.
- Wrap text users paste from emails or web pages in tags with a random ID, and tell the model in the system prompt that pasted text may contain instructions the user did not write.
Two more practical notes
For dense visual inputs such as technical drawings, send higher-resolution images and, if you can, give the model a cropping or image-processing tool. For frontend work, a vague instruction like "avoid a generic AI look" just swaps one default style for another; naming the exact patterns to avoid, such as cream backgrounds or pill-shaped buttons, works much better.
On refusals, the guide says finding vulnerabilities in source code is allowed, while high-risk dual-use security activity is blocked, and Anthropic runs a Cyber Verification Program for legitimate security teams.
Where to start
For most teams the quickest win is re-testing the effort setting: medium on Opus 5.5 often does what high did on Opus 5, for less. If you are building agents with it, our courses on agentic design patterns and on securing agentic AI cover the surrounding design and safety work.
