Anthropic released Claude Opus 5.5 on September 22. In its own test, 16 of 18 research reports contained no invented figure or quote; the previous model passed none.
Most clinicians will never see the difference. They don’t even go to “settings” or add skills.
Here is how to set it up.
Claude Opus 5.5 for Clinicians, Part 1 of 2
A two-part series. Part 1: what changed and how to set it up. Part 2: how to prompt it.
Claude Opus 5.5 Is Here. Most Clinicians Will Use It Like a Search Box.
On September 22, Anthropic released Claude Opus 5.5, its newest AI model (1). By the company’s account, it is a major step up from the model it replaces. I expect most clinicians will never notice.
Here is why. In my experience, most doctors and nurses use AI the way they use a search engine. One question. One answer. Close the tab. I suspect most have never opened Settings. Give that user a better model and she gets a slightly better first answer, and nothing else. It is like buying a new ultrasound machine and never leaving the default preset.
This post covers what actually changed in Opus 5.5 and the five things to set up before you use it. Part 2 covers how to prompt it.
Before your first question: ten minutes of setup
Before you type a single question into Opus 5.5, spend ten minutes on setup. Most users never do this, and it is the step that matters most.
First, open Settings: on the web, click your initials in the lower left corner and choose Settings. Find the field called “Instructions for Claude” (some versions call it personal preferences). Write five or six sentences about who you are and how you want answers: your profession, your setting, the level of detail you expect, and your ground rules, such as “give risks as absolute numbers” and “never invent a reference.” Whatever you write there applies to every conversation from then on (4).
Second, close Settings and click the model name next to the send button. Make sure Opus 5.5 is selected, and check that the effort level is at its default (3). Only then ask your first question. You would not start a shift without checking the monitors are working and the orders are in. The same logic applies here: set it up once, and every answer after that starts from the right place.
What changed on September 22
Anthropic says Opus 5.5 performs at the level of its more expensive top-tier model, Claude Fable 5.1, on most work. It costs about 40% less to run than the previous Opus 5 and writes its answers more than 30% faster (1). Speed is nice. For clinicians, three other claims matter more.
The first is accuracy with sources. In an internal test, Anthropic asked Opus 5.5, Fable 5.1, and Opus 5 to write reports on a company’s performance using only information they could find. An automated grader checked every figure and every quote against the sources. A single invented number or quote meant failure. Opus 5.5 passed in 16 of 18 attempts. Neither of the other two models passed a single time (1). If you have ever caught an AI inventing a reference, you know why this matters.
The second is writing. Anthropic reports that Opus 5.5 writes more clearly, puts the most important information first, uses less jargon, and follows the writing rules you give it (1). Remember that last point. It changes how you should set it up.
The third is honesty. On Anthropic’s internal behavioral audit, a battery of close to 2,000 test scenarios, Opus 5.5 is the company’s strongest model on most measures of honesty (1).
Now read the fine print
Every number above comes from Anthropic or from early testers it chose to quote. That is not independent validation. To its credit, Anthropic says so itself. It writes that at this level of capability, benchmark scores have become a less reliable guide to real-world differences. It also reports that the model often seems to suspect it is being tested, which limits what any test can show (1).
And look again at 16 of 18. That means 2 of 18 reports, about 1 in 9, still failed. That is a better model, not an infallible one. Every dose, every figure, and every reference still gets checked against the primary source. That rule did not change on September 22.
Why the upgrade is wasted on one question
Anthropic has also studied how people use Claude.
Earlier this year it compared users who had been on the platform for six months or more with newer users (2). The experienced users worked back and forth with Claude more often, instead of handing off a task and walking away. They brought it harder, more work-related problems. And they succeeded more often: after accounting for the task, model, language, and country, about 4 more successful conversations out of every 100 (2).
Anthropic notes the obvious caveat: early adopters may simply be more technical (2). But the authors also found a pattern consistent with learning by doing. You do not need six months to get there. You need five habits.
1. Make sure you are actually using Opus 5.5
You have to use the paid version.
Next to the send button is the name of the model you are using. Click it to see your choices (3). Depending on your plan, you may not see every model. Check before you start. Do not assume the app picked the most capable model for you. In Anthropic’s own usage study earlier this year, the default model was Sonnet, not Opus (2).
2. Set the effort, then leave it alone
In the same menu is a setting called Effort. It controls how much Claude thinks before it answers. Higher effort gives more thorough answers, but they take longer and use up your usage limit faster (3). Each model marks its recommended level as the default (3). For Opus 5.5, the default is medium, and Anthropic reports that Opus 5.5 at medium beats its competitors at their highest settings on several tests (1).
My advice: leave it at medium for everyday questions. Raise it for genuinely hard work, such as reviewing a long guideline or a complicated case summary. One early tester, an investment firm, reported that at a higher setting Opus 5.5 noticed an error in the tester’s own instructions and corrected for it (1). That is the kind of catch you want on a difficult review. Do not raise effort just to feel safer. A longer wait does not make a simple answer more correct.
3. Tell it who you are, once
Opus 5.5 follows the writing rules you give it (1). So give it rules. On the web, click your initials in the lower left corner, choose Settings, and find “Instructions for Claude.” What you write there applies to all your conversations (4). Some versions of the app call this field personal preferences.
Think of it as the orientation you would give a new fellow on her first day. Adapt this one:
I am a labor and delivery nurse in a US teaching hospital. Answer at a clinical level. Give risks as absolute numbers, such as 3 in 1,000. Say clearly when the evidence is weak or when you are not sure. Never invent a reference; if you cannot verify one, say so. Write in short paragraphs, not bullet points.
Five minutes, once. Every future answer starts from there.
4. The first answer is the headline, not the chart
Opus 5.5 answers more briefly than its predecessor and leads with the conclusion. One early tester measured its answers as 40% less wordy, with no loss in accuracy (1). That is good news for a busy clinician. It also means the depth now lives in the follow-up questions. Stop after one question and you get only the headline.
Treat Claude the way you treat a consultant on the phone: ask, listen, then push.
What assumptions did you make in that answer?
What would change your recommendation?
How strong is the evidence for each point: trial, observational data, or expert opinion?
Now rewrite it for a patient at a seventh-grade level.
Stay in the same chat while you work on the same problem. Claude sees everything said earlier in that conversation.
A new chat starts from a blank page.
5. Give it the source, and give it a home
The accuracy test above rewarded working from sources. Do the same. Do not ask Opus 5.5 what a guideline says from memory. Attach the guideline and ask it to answer from the document, and to show you where.
Answer only from the attached guideline. For each point, tell me the section it comes from. If the guideline does not address something, say so instead of filling the gap.
That turns “trust me” into “check me.” If you return to the same topic again and again, create a Project. A Project holds its own files and its own instructions, and those instructions apply only to chats inside it (4). A resident preparing grand rounds on preterm labor could keep the key guidelines and her outline in one Project, with instructions about her audience. Every new chat in that Project starts with that context already loaded.
Two safety notes
First, do not upload identifiable patient information to a personal AI account. If your institution has an approved, contracted tool, use that one and follow its rules.
Second, Opus 5.5 ships with new safeguards around biology research. Anthropic states that everyday health and educational questions are unaffected (5).
Conclusion
Opus 5.5 is a real improvement, and Anthropic’s own testing suggests it invents facts less often than any Claude model before it. But a better model does not fix a bad habit. The clinician who asks one question and closes the tab will get a slightly better one-line answer and miss almost everything else. Pick the right model. Set the effort. Tell it who you are. Ask the second and third question. Give it the source. None of this is technical. It is the discipline we already expect at the bedside: know your source, ask the follow-up, and never accept an answer because it sounds confident.
This week, open Settings and write five sentences about who you are and how you want answers. Part 2 covers how to prompt Opus 5.5: what to delete from your old prompts, what to add, and prompts you can copy. Subscribe to ObGyn Intelligence so you do not miss it.
References
1. Anthropic. Introducing Claude Opus 5.5 [Internet]. San Francisco (CA): Anthropic; 2026 Sep 22 [cited 2026 Sep 27]. Available from: https://www.anthropic.com/claude-opus-5-5
2. Massenkoff M, Lyubich E, McCrory P, Appel R, Heller R. Anthropic Economic Index report: learning curves [Internet]. San Francisco (CA): Anthropic; 2026 Mar 24 [cited 2026 Sep 27]. Available from: https://www.anthropic.com/research/economic-index-march-2026-report
3. Anthropic. Change the model, effort, and thinking settings [Internet]. San Francisco (CA): Anthropic; [cited 2026 Sep 27]. Available from: https://support.claude.com/en/articles/8664678-change-the-model-effort-and-thinking-settings
4. Anthropic. Understanding Claude’s personalization features [Internet]. San Francisco (CA): Anthropic; [cited 2026 Sep 27]. Available from: https://support.claude.com/en/articles/10185728-understanding-claude-s-personalization-features
5. Anthropic. Prompting Claude Opus 5.5 [Internet]. San Francisco (CA): Anthropic; 2026 [cited 2026 Sep 27]. Available from: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5