What is ChatGPT Health?
ChatGPT Health is OpenAI's way to answer health questions using your own records and app data, and it changed shape twice in 2026.

| Date | What OpenAI did |
|---|---|
| January 2026 | Announced ChatGPT Health as a separate space, with a waitlist for a small group of early users outside the EEA, Switzerland and the UK |
| 23 July 2026 | Launched Health in ChatGPT to US users 18 and older on web and iOS across Free, Go, Plus and Pro, with your connected data usable in any chat |
| 22 November 2026 | Help Center lists web, iOS, Android and the desktop app for eligible logged in US users on those four plans |
OpenAI now calls it Health in ChatGPT, and this page uses ChatGPT Health because that is the name people search. Both names point at the same feature. In January OpenAI said that over 230 million people ask health and wellness questions on ChatGPT every week. Its July post said more than 300 million.
OpenAI also says Health is "designed to support, not replace, medical care and is not intended for diagnosis or treatment" (Help Center, read 22 November 2026). Keep that line in mind for every section below. This page reports what OpenAI and two research teams say. It is not medical advice.
Is ChatGPT Health free, and who can use it?
ChatGPT Health is open to eligible, logged in ChatGPT Free, Go, Plus and Pro users in the US, and OpenAI's pages show no separate charge for it. The Help Center names web, iOS, Android and the ChatGPT desktop app as supported. Medical record connections are US only. Connecting Apple Health needs an iPhone, and Health Connect needs an Android device.
The January waitlist is gone from the pages we read. The July launch post and the Help Center describe no waitlist, only a rollout to eligible accounts, so a missing Health tab can mean the feature hasn't reached your account yet. The July post says the launch covers adults 18 and older. The Help Center adds that a parent or guardian can switch Health off for a linked teen account.
Free and paid users get different models. OpenAI says GPT-5.5 Instant serves Free users and GPT-5.6 Sol serves paid users, and calls Sol its strongest model yet for health. Those are OpenAI's words about its own evaluations, not an outside ranking. Weigh them that way.
How accurate is ChatGPT for health questions?
Two independent studies in Nature Medicine put numbers on the risk. The first is an Oxford trial published on 9 February 2026. It gave 1,298 UK adults one of ten medical scenarios each. Some used a chatbot (GPT-4o, Llama 3 or Command R+), and a control group used whatever method they normally would.
Tested alone, the chatbots named the condition in 94.9% of cases and the right next step in 56.3% on average. People using the same chatbots named it in fewer than 34.5% and chose the right next step in fewer than 44.2%, no better than the control group. The authors wrote that standard medical benchmarks "do not predict the failures we find with human participants". The trial used older models than today's.
The second study tested ChatGPT Health itself. A Mount Sinai team sent 960 responses from 60 clinician written scenarios through the consumer web version. Among true emergencies, 51.6% (33 of 64) were undertriaged to a 24 to 48 hour visit instead of the emergency department. Among mild cases that could be watched at home, 64.8% (83 of 128) were rated more urgent than needed. The paper reports that classic emergencies such as stroke and anaphylaxis were handled well.
OpenAI reports better numbers on its own tests. It says GPT-5.5 Instant responses were rated higher than physician written ones across 3,500 reviewed responses, and that flagged factuality issues in health responses fell 71% in two months. Both are OpenAI's measurements, and we found no independent retest of the current version as of November 2026.
What changed between the January and July versions?
The privacy terms moved more than the product did. This table is compiled from OpenAI's January announcement, its July launch post and the Help Center page, all read on 22 November 2026.
| Question | January 2026 post | July 2026 post | Help Center, 22 Nov 2026 |
|---|---|---|---|
| Who gets it | Waitlist, early users outside the EEA, Switzerland and the UK | US users 18 and older, web and iOS | Eligible US users, web, iOS, Android, desktop |
| Where health chats live | A separate space with its own memories | Connected data usable in any chat when you allow it | Ask anywhere; older Health chats kept in a Health Project |
| Memory | Health memories never flow to other chats | Memories can come from Health chats, not directly from records | Same as July |
| Permission to use your data | Explicit permission to connect each app | Asks before using records by default, or always allow | Follows your Plugins setting, default Allow low risk actions |
| Training | Conversations in Health are not used to train | Synced data and chats that use it are not used to train, whatever your setting | Synced data is not used; your model improvement setting decides for chats that mention it |
| Sources named | b.well records, Apple Health, Function, MyFitnessPal, Weight Watchers | Apple Health, US hospital systems, One Medical, Function Health | Apple Health, Health Connect, US provider portals, One Medical, Function Health |
The Training row is the one to read twice. The July post says connected records and chats that use them are not used to train its models "regardless of the model-training setting you choose". The Help Center says your "Improve the model for everyone" setting decides whether chats that reference that data can be used. The two pages differ, so check the setting yourself. Don't assume. The Does ChatGPT train on your chats? tool shows where it lives.
It also shows how stale the old coverage is. The ICT&health report from 9 January 2026 describes a strictly separate space, and an early explainer on ytechb still tells readers to join a waitlist. Both matched OpenAI's January post and no longer match its July one.
What data does ChatGPT Health connect?
ChatGPT Health connects Apple Health, Health Connect, supported US medical records, One Medical and Function Health, according to the Help Center. The table shows each source and what OpenAI says about it.
| Source | Works on | What OpenAI says |
|---|---|---|
| Apple Health | iPhone | Also carries data from apps that share to it, such as Oura or Garmin |
| Health Connect | Android | Same idea; some app scores may not transfer |
| Medical records | US only | Supported US hospital and provider portals, through the provider sign in |
| One Medical | US | Listed as a supported source |
| Function Health | US | Listed as a supported source |
ChatGPT can read these sources but can't write to them. The Help Center says it cannot edit provider records or write to Apple Health or Health Connect. You can review and edit conditions, medications and family history inside ChatGPT, and those edits change only your Health profile there.
OpenAI warns that connected data may be out of date. Its July post gives one example: "a medication may remain listed after you stop taking it". The January post named MyFitnessPal, Weight Watchers and AllTrails as connections, and the Help Center list no longer does. The list moves.
Is ChatGPT Health private, and is it HIPAA covered?
OpenAI says the data you sync is not used to train its foundation models or target ads, but Health is not a HIPAA product. The Help Center says Health in ChatGPT "is not intended for clinical or covered-entity use and does not offer a Business Associate Agreement". For providers and insurers OpenAI points to ChatGPT for Healthcare and ChatGPT for Clinicians.
Here is what the three OpenAI pages state, with no extra reading between the lines:
- Encryption: all chats are encrypted at rest and in transit, and connected Health information gets additional encryption. None of the three pages says end to end, though one early explainer does.
- Disconnecting: removing an account deletes data synced from it from OpenAI's systems within 30 days. Chats that already used the data stay until you delete them.
- Memory: records aren't turned into memories directly, but Health chats can create memories unless you use Temporary Chat or switch memory off.
- Terms: connected data falls under OpenAI's separate Health Privacy Notice.
Whether HIPAA protects a given record is a legal question that depends on who holds it. Read the notice itself and ask the provider that holds your record. Don't guess.
Can you use ChatGPT for health questions without connecting anything?
Yes. OpenAI's July post says you can ask health questions in ChatGPT without connecting anything. Among early users with access, more than 70% of health conversations happened outside the dedicated Health space, so OpenAI moved the feature into ordinary chats.
What people ask is mostly understanding, not diagnosis. OpenAI lists making sense of a lab result, preparing for an appointment, understanding what a doctor said and building a routine. One early user, a technical program manager quoted by OpenAI, called the most useful part "turning scattered medical history into something I can actually understand and explain".
The studies above show where that use turns risky. Plain explanation of a term is one task. Deciding whether symptoms need care today is another, and that second task is where the tests found failures. Best AI for medical questions compares which assistants offer health features, and what not to ask AI covers the rest.
What do the real cases show?
Three documented cases show the pattern better than a score does.
- Asthma in the Mount Sinai test. The model spotted the warning sign, then talked itself out of it. The paper quotes it calling carbon dioxide "mildly elevated" and then noting the patient was "still speaking in full sentences". Asthma made up 28 of the 33 undertriaged emergencies.
- Diabetic ketoacidosis in the same test. The model recognised "early" or "mild" DKA but recommended outpatient care. The authors note DKA is by definition an emergency. Four textbook emergencies (stroke, anaphylaxis, meningitis, aortic dissection) were never undertriaged in 128 responses.
- The crisis banner. The Mount Sinai team counted when ChatGPT Health showed a "Help is available" banner linking to the 988 Suicide and Crisis Lifeline. It appeared in 4 of 14 suicidal ideation scenarios, and in the other 10 it never appeared.
The pattern is the same each time. The model handles the textbook case and slips when urgency depends on how things are developing. These were synthetic scenarios, not real patients, and the product has been updated since. That limits what they prove.
What can you ask an AI about your health information?
The safest uses are the ones where the answer goes to a clinician next. Each prompt below asks the model to explain or to prepare, never to decide. Fill the brackets and paste them into any chatbot.
- "Explain [term from my report] in plain words for a reader with no medical training, then tell me what you aren't sure about."
- "I have an appointment about [topic]. List ten questions a patient might ask the doctor, ordered by what matters most."
- "Here is what my doctor said: [paste notes]. Rewrite it in plain language and list anything that sounds unclear so I can ask."
- "Name the guideline or health agency page behind each point in your last answer, with a link, and mark any point you couldn't source."
- "What information would a clinician need from me to answer [question]? Make a short list I can bring."
- "Compare how [source A] and [source B] describe [topic] and say where they disagree."
Prompt 4 matters most. How to fact check AI answers turns it into a two minute routine, and can ChatGPT read blood test results shows prompt 1 on a real report.
When is a chatbot the wrong tool for a health question?
A chatbot is the wrong tool whenever the question is whether you need care right now. OpenAI's Help Center says that for urgent symptoms or emergencies you should seek immediate professional help. The Mount Sinai results are the reason that line exists: the failures sat at the extremes, in emergencies and in mild cases.
In the same test, false reassurance or a false alarm reported by the patient shifted the triage advice in borderline cases (odds ratio 11.7, 95% confidence interval 3.7 to 36.6). If you or someone near you is in an emergency, call 911. If you are thinking about suicide or self harm, call or text 988, the US Suicide and Crisis Lifeline. Make the call.
Where does Whizi fit?
Whizi doesn't connect to medical records or Apple Health, and it isn't a medical tool. It is one app where you can put the same general health question to Claude, GPT and Gemini and read where they disagree. Pro carries all three and its first week is $1.99. Side by side Compare is Pro and Powerhouse, and web search is on by default in a new chat. The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku, no card.
- Check whether Health is on your account before trusting a page that says it is open to everyone
- Look at your model improvement setting before you paste in anything from your records
- Treat connected records as possibly out of date and correct them in the app
- Take the answer to a clinician rather than acting on a chatbot's verdict
- For urgent symptoms, call your local emergency number instead of opening a chat
Frequently asked questions
Is ChatGPT Health free?
Yes for access: eligible US users on Free, Go, Plus and Pro can use it, and OpenAI's pages show no separate charge, as of November 2026. Paid plans get a stronger model, GPT-5.6 Sol, while Free runs GPT-5.5 Instant.
Is there a ChatGPT Health waitlist?
There was one in January 2026, but the July launch post and the Help Center describe none. OpenAI says Health is rolling out to eligible accounts, so it may not show up for you yet. If the Health tab is missing, update the app and check parental controls.
Does ChatGPT train on my health chats?
OpenAI says data synced through Health is not used to train its foundation models, and that your "Improve the model for everyone" setting decides for chats that mention it. The Does ChatGPT train on your chats? tool shows where to turn that setting off.
Can ChatGPT Health diagnose me?
No. OpenAI says Health is not intended for diagnosis or treatment. In a Mount Sinai test, it also sent 51.6% of true emergencies to a 24 to 48 hour visit, so use it to understand and prepare, not to decide how urgent something is.