Skip to content
TBSI.

Start here

  • LLM LeaderboardCompare frontier models on intelligence, price and speedTool
  • What is the best superintelligence?The best frontier model for each jobModels
  • Frontier Gap ReportHow far apart the leading labs areReport
  • MethodologyHow we verify sources, data and claimsAbout

Best AI for Medical Questions (2026): ChatGPT Health, Claude and OpenEvidence

Which AI to use for understanding a diagnosis, reading lab results or preparing for an appointment, which one clinicians use, and when not to ask an AI at all

By TheBestSuperintelligence Editorial TeamUpdated

◈ Executive SummaryCited · Documentation

What is the best AI for medical questions?

For most people in the US, the best AI for medical questions is ChatGPT with its Health space: it can read records you connect, keeps them out of model training, and works on the Free plan. Clinicians should use OpenEvidence. Neither replaces a doctor.

☰On this page (8 sections)
  1. 01.Which AI should you use for a medical question?
  2. 02.ChatGPT Health: best AI for health questions about your own records
  3. 03.Claude: the alternative if you already pay for it
  4. 04.OpenEvidence: best AI for doctors
  5. 05.Is AI medical advice safe?
  6. 06.How to ask AI a medical question safely
  7. 07.Limits of this guide
  8. 08.Sources & Further Reading

Which AI should you use for a medical question?

The best AI for medical questions, for most people in the US, is ChatGPT used inside its Health space. The deciding reason is the setup around the model rather than the model alone: Health keeps health conversations, connected apps and files separate from your other chats, can read medical records and Apple Health data you connect, and OpenAI says that data is not used to train its foundation models or target ads. It runs on the Free plan. If you are a clinician, use OpenEvidence instead. Whatever you use, an AI is not a doctor, and nothing on this page is medical advice.

Your job Pick Why
Understand a lab result, diagnosis or visit note ChatGPT (Health) Reads records you connect; separate, untrained-on space
Same, if you already pay for Claude Pro or Max Claude US record connectors in beta; no training on health data
Prepare questions for an appointment Either Both vendors name this as a core use
Look up evidence for patient care OpenEvidence Content agreements with NEJM, JAMA, Cochrane and others
Chest pain, stroke signs, trouble breathing No AI Call emergency services or go to the emergency department

Our take: the gap between the big general assistants on everyday health questions is smaller than the gap between using one carefully and using one carelessly. The how-to section below matters more than the pick.

ChatGPT Health: best AI for health questions about your own records

OpenAI launched ChatGPT Health on January 7, 2026, and on July 23, 2026 began rolling Health out to logged-in US users aged 18 and older on web and iOS, across the Free, Go, Plus and Pro plans. OpenAI says more than 300 million people turn to ChatGPT with health-related questions every week.

What it does that a plain chat does not:

  • Your records as context. You can connect medical records (through OpenAI's partner b.well), Apple Health (iOS only), and apps such as Function and MyFitnessPal. OpenAI's examples: comparing a new result with prior tests, summarising changes since your last appointment, explaining a visit note in plain language.
  • A separate space. Health has its own memories. OpenAI says Health information never flows back into your other chats, and those chats cannot access files or memories created in Health.
  • Physician input. OpenAI says it worked with more than 260 physicians who have practised in 60 countries, who gave feedback on model outputs over 600,000 times, and that it evaluates the model with HealthBench, its own rubric-based test.

The limits come from OpenAI too. It says Health "is not intended for diagnosis or treatment" and is designed to support, not replace, medical care. Medical record integrations are US-only. In July 2026 OpenAI called GPT-5.6 Sol, for paid users, "our strongest model yet for health"; we could not confirm whether its newer GPT-6 models changed that. HealthBench is OpenAI's own evaluation, so treat its scores as vendor claims. For the wider model field, see the most advanced AI right now.

Claude: the alternative if you already pay for it

Anthropic announced its healthcare push on January 11, 2026. Its page says that "in the US, Claude Pro and Max plan subscribers can choose to give Claude secure access" to their lab results and health records, through HealthEx and Function connectors launched in beta, with Apple Health and Android Health Connect integrations rolling out in beta in its mobile apps.

On privacy, Anthropic states: "We do not use users' health data to train models." Sharing is opt-in, you choose what to share, and you can disconnect or change permissions at any time. Anthropic says Claude is designed to include contextual disclaimers, acknowledge uncertainty and direct users to healthcare professionals for personalised guidance.

Two caveats. The record connectors require a paid plan, unlike ChatGPT Health. And Anthropic's "HIPAA-ready" products are, in its own words, specific to healthcare providers and payers; they do not describe the consumer integrations.

Our take: if you already use Claude for work and pay for Pro or Max, there is no strong reason to switch for health questions. If you are choosing from scratch on a free plan, ChatGPT Health offers more.

OpenEvidence: best AI for doctors

OpenEvidence describes itself as a medical platform for healthcare professionals, with answers grounded in peer-reviewed research. Its site lists content agreements with NEJM, JAMA, Nature, NCCN and Cochrane, collaborations with societies including the AAP, ACC, ACOG and ACEP, and says it is free for verified US healthcare professionals.

That makes it a different kind of tool from a consumer chatbot: it is built for a clinician who can judge the evidence it cites. Anthropic's usage policy points the same way for anyone building on Claude: when a medical recommendation such as a diagnosis or treatment advice is a high-risk recommendation, "a qualified person must meaningfully review" it before it reaches the patient, and the patient must be "clearly told that AI was used to produce it."

Is AI medical advice safe?

It is safe enough for understanding and preparation, and not safe as a substitute for diagnosis. Three sources set the boundary.

Real users do worse than benchmarks suggest. An Oxford-led randomised study, which the University of Oxford reported as published in Nature Medicine on February 10, 2026, gave 1,298 participants doctor-written medical scenarios. Tested alone, the models (GPT-4o, Llama 3 and Command R+) identified conditions in 94.9% of cases. Participants using the same models identified relevant conditions in under 34.5% of cases and the right course of action in under 44.2%, no better than a control group. The researchers found that users did not know what to tell the model, that slight changes in wording produced very different answers, and that replies mixed good and bad advice. The models tested are older than today's, so the exact figures may not carry over; the failure modes probably do.

WHO names the risks. WHO's January 2024 guidance on large multi-modal models, with over 40 recommendations, warns that they can produce "false, inaccurate, biased, or incomplete statements", that patients and clinicians can fall into "automation bias", and that the systems are "vulnerable to cybersecurity risks".

A chatbot is not a cleared medical device. The FDA keeps a list of AI-enabled medical devices authorised for marketing in the US. It says the list is not comprehensive, and that it plans to explore ways to tag devices that incorporate foundation models. OpenAI, for its part, says Health is not intended for diagnosis or treatment.

On privacy, the vendor commitments above are promises in each company's own policies. Read the connection screen before linking a record, and disconnect apps you no longer use.

How to ask AI a medical question safely

The Oxford study's failure modes suggest the fixes. To ask AI a medical question well:

  1. Give the full picture. Age, sex, existing conditions, current medications and doses, how long symptoms have lasted, and what has changed. Users in the study often left out what mattered.
  2. Ask it to explain, not to decide. "What does a raised ALT on this report usually mean, and what would a doctor check next?" is safer than "What do I have?"
  3. Ask what would make it urgent. Request the warning signs that mean you should seek care today, and act on them without waiting for a second answer.
  4. Ask twice, differently. Wording changed answers in the study. If two phrasings disagree, that is a question for your clinician.
  5. Ask for sources and check one. Prefer answers that point to a health body or a guideline you can open.
  6. Turn it into questions for your doctor. End with: "List five questions I should ask at my appointment." Both OpenAI and Anthropic name appointment preparation as a core use.
  7. Share the minimum. Leave out names, addresses and ID numbers unless a feature needs them.

For long reports or discharge letters, the prompts in our guide to summarising documents with AI help you get a plain-language summary to take to your appointment.

Limits of this guide

  • Availability is US-centred: we could not confirm when ChatGPT Health or Claude's record connectors reach other countries.
  • The Oxford figures come from 2024-era models; we did not find a comparable study of current ones.
  • We did not review health features in Google's Gemini or Microsoft's Copilot.
  • Vendor health evaluations, including HealthBench, are self-reported.

Sources & Further Reading

Claim Source Date
ChatGPT Health launch, separate space and memories, connected apps, US-only records, 260+ physicians in 60 countries, 600,000+ feedback instances, "not intended for diagnosis or treatment" OpenAI: Introducing ChatGPT Health (read via Internet Archive copy of 2026-11-06) Jan 7, 2026; read 2026-11-17
US rollout to users 18+ on web and iOS, Free/Go/Plus/Pro; 300M+ weekly health askers; records not used for training or ads; GPT-5.6 Sol "strongest model yet for health" OpenAI: Launching Health in ChatGPT (read via Internet Archive copy of 2026-11-16) Jul 23, 2026; read 2026-11-17
Claude Pro and Max US record access, HealthEx and Function connectors in beta, no training on health data, disclaimers, HIPAA-ready scope Anthropic: Advancing Claude in healthcare and the life sciences Jan 11, 2026 (edited Feb 7, 2026); read 2026-11-17
HealthEx connector connects Claude to health records Claude: HealthEx connector Read 2026-11-17
High-risk medical recommendations need qualified human review and disclosure of AI use Anthropic Usage Policy Effective Nov 12, 2026; read 2026-11-17
OpenEvidence audience, content and society partners, free for verified US healthcare professionals OpenEvidence (read via Internet Archive copy of 2026-11-16) Read 2026-11-17
Oxford-led user study: 1,298 participants, 94.9% vs under 34.5% and 44.2%, three failure modes Bean et al., Clinical knowledge in LLMs does not translate to human interactions (arXiv) Apr 26, 2025; read 2026-11-17
Study published in Nature Medicine; LLMs no better than traditional methods; quotes from the authors University of Oxford news release (read via Internet Archive copy of 2026-11-13) Feb 10, 2026; read 2026-11-17
WHO guidance on large multi-modal models: 40+ recommendations, named risks WHO news release Jan 18, 2024; read 2026-11-17
FDA list of AI-enabled medical devices is not comprehensive; plans to tag foundation-model devices FDA: AI-Enabled Medical Devices list Read 2026-11-17

Frequently asked questions

What is the best AI for medical questions?

For most people in the US, the best AI for medical questions is ChatGPT with its Health space: it can read records you connect, keeps them out of model training, and works on the Free plan. Clinicians should use OpenEvidence. Neither replaces a doctor.

Can I use OpenEvidence if I am not a doctor?

OpenEvidence describes itself as a platform for healthcare professionals, and its site says it is free for verified U.S. healthcare professionals. Patients are better served by a general assistant used to understand information and prepare questions, then checking the answer with their own clinician.

Does ChatGPT use my health records to train its models?

OpenAI says connected medical records, Apple Health information and the conversations that use them are not used to train its foundation models or to target ads. You can remove medical record access in the Apps section of Settings, and a disconnected app loses access immediately.

Is ChatGPT Health available outside the United States?

OpenAI's July 23, 2026 update describes a rollout to logged-in users aged 18 and older in the U.S., on web and iOS, across Free, Go, Plus and Pro plans. Medical record integrations were U.S.-only from launch. We could not confirm a date for other countries.

Do AI chatbots give better medical advice than a Google search?

Not in the largest user study so far. In an Oxford-led randomised trial with 1,298 participants, people using GPT-4o, Llama 3 or Command R+ identified relevant conditions in under 34.5% of cases, no better than a control group using sources of their choice.

How we sourced this. We read the vendors' health pages and policies, WHO's guidance and the Oxford study on 2026-11-17. We did not test the products with real health data. Methodology