How to prompt Claude Opus 5.5: what to change (Nov 2026)
Anthropic's Opus 5.5 prompting guide in plain English: why old habits waste tokens, the new medium effort default, and the prompts to drop or add.
Checked and updated 15 November 2026 at 03:50 UTC

Key points
- Anthropic's official guide says prompts written for Claude Opus 5 still work on Opus 5.5, but some old habits now cost time and money, such as telling the model to "think carefully" or to write out its reasoning.
- The biggest change is effort. Opus 5.5 defaults to medium (Opus 5 defaulted to high), thinks more per turn at the same level, and at medium matched or beat Opus 5 at high in Anthropic's tests.
- Specific instructions work better than vague ones. Name the design styles to avoid, tell agents to look through your apps before acting, and mark text you pasted from elsewhere.
To get the best out of Claude Opus 5.5, give it a clear goal, pick the effort level on purpose (the default is now medium), and delete old prompt lines that tell it to "think carefully" or write out its reasoning. That's the short version of Anthropic's official Prompting Claude Opus 5.5 guide, which is getting fresh attention this week as people find their old prompts make the model slower and pricier than it needs to be.
The good news: Anthropic says prompts written for Claude Opus 5 "should perform well without changes." The tweaks below are about speed, cost and a few specific behaviors.
What changed in Claude Opus 5.5
Anthropic released Opus 5.5 on September 22, 2026. Compared with Opus 5, the docs say it:
- Writes output more than 30% faster and usually finishes the same task with fewer tokens
- Always thinks before answering. Thinking can't be switched off; you control how much with the effort setting
- Defaults to medium effort (Opus 5 defaulted to high)
- Thinks more per turn at the same effort level, most of all at xhigh and max
- Reads charts, diagrams and screenshots much more accurately, and is better at computer use
It's also cheaper on the API: $4 per million input tokens and $20 per million output tokens, versus $5 and $25 for Opus 5 (What's new in Opus 5.5).
Old habits to drop
| Old habit | Why it hurts on Opus 5.5 | What to do instead |
|---|---|---|
| Running at high or max effort "to be safe" | Opus 5.5 thinks more per level, so turns run longer and cost more | Start at medium; in Anthropic's tests it matched or beat Opus 5 at high |
| "Think carefully before you answer" in a chat system prompt | The model already decides how much to think; Anthropic found removing the line made replies start sooner with no clear quality loss | Delete it and use the effort level |
| "Write out your reasoning step by step in your answer" | It can be declined as reasoning extraction | Ask for a short explanation or summary instead |
| Prompting less thinking with words | Prompt instructions are less reliable | Lower the effort level first |
| "Avoid a generic AI look" for web designs | It just swaps one default style for another | List the exact patterns to avoid |
New habits that help
Be specific about what you don't want. For websites and apps, Anthropic's own example bans cream backgrounds, italic accent words in headlines, "01/02/03" section labels, monospace labels and pill-shaped buttons. Check what the first draft used and extend the list.
Tell agents to look around first. If Claude works across email, docs, spreadsheets and other apps, one sentence asking it to open everything that could be relevant before acting helped it finish noticeably more tasks correctly in Anthropic's testing.
Mark pasted text. When you paste an email or web page, wrap it in tags so Claude knows those words aren't your instructions. Anthropic says Opus 5.5 resists hidden instructions in pasted content best when you do this.
Settle earlier answers. In long chats, Opus 5.5 sometimes re-checks an earlier answer. If you'd rather it didn't, say that answered questions are done unless you raise them again. Skip this for long analyses where you want it to keep double-checking.
For people building with the API
thinking: disablednow returns an error. Use effort instead:lowis the replacement for thinking-off setups.- Forced tool use (
tool_choiceany or tool) is not supported; useautowith strict tool use. - Set
max_tokenshigh enough for thinking plus the reply. Anthropic says 128,000, the maximum, has worked well for long coding turns. - Unattended agents may stop after a progress update. Keep a task checklist and nudge it to continue, but cap automatic nudges at two or three.
For more on how Anthropic's models compare, see our Claude Fable 5.2 explainer, and if Claude stops responding, check is Claude down.
Frequently asked questions
Do I need to rewrite my prompts for Claude Opus 5.5?
Mostly no. Anthropic says existing Claude Opus 5 prompts should perform well without changes. The main things to revisit are the effort setting, any "think step by step" or "show your reasoning" lines, and agent setups that treat a progress update as the end of a task.
Why does Claude Opus 5.5 seem slower or use more tokens than Opus 5?
At the same effort level, Opus 5.5 tends to think more per turn than Opus 5, especially at xhigh and max. If you carried over an old high or max setting, Anthropic suggests starting at medium (the new default) and lowering effort before trying prompt tricks.
Is Claude Opus 5.5 free?
Not on the Free plan. Notebookcheck reported at launch that Claude's pricing page lists Opus 5.5 for Pro, Max and Team, while Free stays on Sonnet and Haiku. On the API it costs $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5.
Sources
Related
Claude for Teachers: free until July 30, 2028 (how to sign up)
Anthropic extended free Claude for Teachers for verified US K-12 educators, schools and districts through July 30, 2028. What's included and how to join.
What is Lean? The free proof checker behind the Riemann proof
Lean is a free, open-source language where a computer checks every step of a math proof. It just verified the Claude-made Riemann hypothesis proof.
Riemann hypothesis proof with Claude: Lean certifies it (Nov 2026)
Yes, the computer check passed: on Nov 21 the Lean and Mathlib communities certified Levent Alpöge's Claude-made Riemann hypothesis proof. Human review is next.