OpenAI's healthcare strategy combines patient-facing health conversations, a product for individual clinicians and institutionally governed software. Its significance lies in connecting those starting points with medical evidence and patient context, not in establishing that one chatbot can safely perform every clinical task. This account is updated to 15 October 2026 and distinguishes published functionality, company-reported evaluations and strategic interpretation.
The January-to-September expansion established a consumer, professional and enterprise product structure. October developments at competitors now sharpen the question: can OpenAI become the interface through which clinicians understand information, while other companies retain the records, clinical relationships and authority to act? This article is published by iatroX and includes its reference and learning products in the comparison; it does not rank products by incomparable benchmark scores.
The three-tier stack
Consumer: ChatGPT Health
OpenAI introduced ChatGPT Health on 7 January 2026, initially with limited access. The announcement described connecting supported US medical records, Apple Health and services including MyFitnessPal, Peloton, Function, AllTrails, Instacart and Weight Watchers. These were the named connections in that announcement, not a claim that every integration has identical availability or data coverage today. The proposed uses included understanding laboratory information, preparing for appointments and considering everyday health questions. OpenAI's January announcement describes a supporting role, not a replacement for professional care.
The development figures also need their original meaning. In January 2026, OpenAI described more than 260 physician collaborators across 60 countries and more than 600,000 feedback interactions over two years. Those are development inputs, not patients enrolled in an outcomes trial. Its separate January healthcare report reported more than 40 million people turning to ChatGPT daily for health questions. The earlier wording on this page, "40 million daily US health queries", conflated people with queries and added an unsupported US-only scope; it should not be used. The figure does not describe adoption of the subsequently launched Health feature.
The interface changed after the initial release. On 23 July 2026, OpenAI announced rollout to logged-in US users aged 18 and over on web and iOS, allowing relevant connected health information to inform conversations outside the original dedicated space with the user's permission. January's description of an isolated health compartment therefore should not be treated as an unchanged account of the July product.
OpenAI's July privacy description says connected medical records, Apple Health information and conversations using that information are not used for foundation-model training or advertising targeting. It also distinguishes those conversations from ordinary chats governed by standard settings. Encryption, permission controls and separation of purposes are relevant safeguards, but not proof of clinical correctness. January exclusions for the UK, EEA and Switzerland and July's US rollout concern this specific health experience, not a prohibition on asking health questions in ordinary ChatGPT elsewhere.
Individual clinicians: ChatGPT for Clinicians
OpenAI announced this professional product on 22 April 2026, free for eligible verified US physicians, nurse practitioners, physician assistants and pharmacists. Its public offer includes clinical research, documentation assistance and repeatable workflow skills. OpenAI's podcast description, checked on 15 October 2026, identifies Nate Gross as Head of Health; leadership provides context but does not establish product effectiveness. The launch announcement and current eligibility documentation, checked on 15 October 2026, should be read together.
Verification uses applicable US professional credentials, including NPI information. That is not the same as saying that no clinician based in the UK could possess US credentials. Equally, possessing an account does not establish permission to use it for every professional purpose in another country.
The product offers optional HIPAA support through a Business Associate Agreement for eligible accounts, according to the current documentation. A free account should not be assumed to have every contractual or organisational permission needed to process identifiable patient information. OpenAI also describes eligible continuing-medical-education activity; this is a specific offering to check against its terms, not a reason to assume every AI conversation earns credit.
Healthcare organisations: ChatGPT for Healthcare
The institutional product was announced on 8 January 2026. OpenAI described organisational controls, support for HIPAA compliance, Business Associate Agreements and customer-managed encryption keys, alongside models from the GPT-5 family and evaluations including HealthBench and GDPval. These are different kinds of assurance: contractual and security arrangements do not turn a model benchmark into a clinical-outcome study. OpenAI's enterprise announcement provides the starting specification.
That January announcement named AdventHealth, HCA Healthcare, Boston Children's Hospital, Cedars-Sinai, Memorial Sloan Kettering, Stanford Medicine Children's Health and UCSF among organisations beginning rollout or working with the offering. Their inclusion does not establish identical implementation, access or clinical deployment across every department. Enterprise purchasing and local adoption remain separate from an individual clinician's free account.
What changed on 1 September 2026
The September announcement brought an authorised connection to Epic information and a separate Healthcare Public Data plugin. The Epic documentation, checked on 15 October 2026, describes a read-only connection constrained by existing access permissions. It can retrieve relevant clinical information; it cannot thereby amend the chart, place an order or send a patient message. UCSF Health was named as the pilot organisation in OpenAI's announcement.
The described information includes notes, medications, conditions, encounters and laboratory results, with FHIR-based access to relevant observations, allergies, diagnostic reports and documents. Actual availability depends on configuration and authorisation. Opening in the right record is not proof that every clinically important fact is available or current.
The Healthcare Public Data plugin connects to nine named sources: PubMed, ClinicalTrials.gov, DailyMed, RxNorm, openFDA, CMS Coverage, CMS Open Data, Medicare Care Compare and the NPI Registry. They answer different questions about literature, trials, medicines, coverage, services and professionals. A source-by-source assessment is more useful than calling the collection a single medical knowledge database.
Eligible ChatGPT for Clinicians accounts can have access to the public-data capability without the enterprise Epic connection. Likewise, a patient connecting their own records through Health is not using the same permission arrangement as a clinician accessing an organisation's Epic environment. Those distinctions matter even where both experiences are described informally as "ChatGPT reading medical records".
GPT-6 Astra: the new foundation model underneath
The original Astra discussion remains relevant as a dated model-development milestone, rather than a declaration of the current model powering every healthcare product. In its 3 September 2026 Astra announcement, OpenAI reported a length-adjusted HealthBench Professional score of 63.4, compared with 60.5 for GPT-5.6 Sol, alongside improved results on the harder evaluation. Those are benchmark scores under specified conditions, not percentages of patients correctly diagnosed.
The announcement described approximately 1.05 million tokens of context and capabilities including web and file search, code execution, computer use, MCP and structured outputs. Such tools could support multistep work, but their availability at model level does not grant access to a hospital record or authorise clinical actions. Nor does a longer context window demonstrate reliable interpretation of a long, contradictory patient history.
OpenAI subsequently published GPT-6.1 Sol documentation on 29 September 2026. Astra should therefore not be labelled OpenAI's latest model in an October update. The model used in a particular healthcare deployment must come from that product's current documentation, rather than inference from the most recent general launch.
The substantive strategic point survives these qualifications. Combining retrieval, reasoning and tools may make more complicated assistance feasible. The evidence required to rely on that assistance still concerns the configured workflow, its users and its consequences. iatroX's dedicated Astra coverage provides the separate model discussion; the product and model timelines should not be collapsed.
OpenAI reports physicians evaluated 27 EHR-related use cases
OpenAI's 1 September 2026 account describes 27 EHR-related use cases, including pre-visit review, timelines, medication review and handovers. It reports 4,363 physician ratings and says 99.1% of responses were rated safe. Importantly, 4,363 is the reported number of ratings, not a verified count of distinct responses or patients. The same announcement reports more than 93% of responses rated good or better for accuracy across each of five tested connected sources, not across all nine available sources. The published account supplies those units and boundaries.
A useful interpretation asks what the raters saw, how cases were selected, which harms the rubric could recognise and whether errors would be detectable after the system entered routine care. A response judged safe in an evaluation is not evidence that the subsequent clinical action improved an outcome. These distinctions do not make the testing irrelevant; they define the conclusion it can support.
For example, in a fictional pre-visit task, an accurate medication summary could still omit a recent change recorded elsewhere. The summary's wording, completeness of the retrieved context, clinician review and eventual management are separate evaluation targets. The example illustrates a question to test, not an incident reported in OpenAI's evaluation.
The technical foundation and its limits
Two benchmark descriptions in the earlier edition require correction. The claims of "100% diagnostic accuracy" and "98.7% on healthcare-specific tests" are not supported by the cited GPT-5.2 material. OpenAI's 11 December 2025 GPT-5.2 announcement places 98.7% under Tau2-bench Telecom, a tool-calling benchmark, not healthcare. It also reports 100% on AIME 2025, a mathematics assessment; neither number establishes diagnostic accuracy. The January physician-feedback programme is a further, separate development claim.
Independent evidence must also be reported correctly. The Nature Medicine study published on 23 February 2026 examined ChatGPT Health using 60 clinical vignettes across 21 domains, varied across 16 experimental conditions to produce 960 responses. It was a structured triage experiment, not a trial of care delivered to 960 patients.
For clear emergency presentations, the reported undertriage rate was 51.6%, rounded to 52%. For non-urgent presentations, accuracy was 35.2% and overtriage was 64.8%. The earlier page's "35% overtriage" wording inverted the relevant finding. The study also reported sensitivity to social framing, with an odds ratio of 11.7 and a 95% confidence interval of 3.7 to 36.6, and inconsistent crisis-resource activation. Framing-related shifts were not all clinically inappropriate.
These findings support scrutiny of the tested triage behaviour. They do not directly replicate or refute the September EHR evaluation: the product configuration, task, inputs and endpoints differ. Nor should February results be silently assigned to Astra or a later release. The appropriate conclusion is to require evidence matched to the proposed use, rather than comparing an emergency-undertriage percentage with a response-safety percentage as though they measure the same thing.
The competitor landscape
OpenEvidence: the valuation and partnership picture has changed
OpenEvidence's 21 January 2026 funding announcement reported a $250 million Series D at a $12 billion valuation, led by Thrive Capital and DST Global. Subsequent September reporting described another $250 million financing at a $15 billion valuation, identifying Andreessen Horowitz and Byers Capital as leads. The January valuation is historical context, not the current reported funding valuation.
Activity and commercial figures require similar discipline. The older estimate of approximately 15 million monthly "clinical consultations" should not be read as a current count of visits delivered by OpenEvidence. Its January announcement instead reported about 18 million platform-supported consultations during December 2025, using the company's terminology. Such activity is not a demonstrated number of unique patients, independent clinical decisions or improved outcomes. The older quoted advertising range of $70 to $150-plus per thousand impressions is not an independently verified current vendor rate card and is not used here to estimate revenue.
The company described an ad-supported service free to doctors and proprietary medically specialised models in January 2026. Its own security statements, including HIPAA-related arrangements and SOC 2 Type II assurance, concern specified controls rather than the clinical correctness of every answer. Buyers should check current scope through the company's security materials, not treat a security claim as a medical-device certificate.
Institutional integration is now a concrete part of the competitive picture. Mount Sinai's 31 March 2026 announcement describes access within its electronic record. Sutter's 11 February 2026 collaboration and Cedars-Sinai's 20 May announcement illustrate additional routes into institutional practice. A collaboration announcement should not be enlarged into uniform deployment or independently demonstrated benefit throughout a health system.
The 13 October 2026 Johns Hopkins Medicine collaboration adds a different dimension: the stated ambition spans clinical care and medical education. That suggests competition over how clinicians learn and judge evidence as well as how they retrieve it. It does not establish a delivered curriculum, a measured educational advantage or equivalence to iatroX's own learning tools.
EHR vendors and specialist products are not standing still
As documented on 15 October 2026, Epic's AI portfolio includes native charting and its Art environment, including patient-information questions, citations and visualisations through Ask Art. Epic is therefore not merely a passive database waiting for another company's assistant. A hospital could adopt native functions alongside an external knowledge service, depending on what each workflow actually requires.
The older description of Medwise as enterprise-only was too categorical. Its current site describes clinical AI access alongside organisational use, while its public procurement listing describes integration of local and national knowledge. The relevant questions are which product is available, what local material it can use and what an organisation must configure, not whether the brand belongs permanently in one purchasing category.
Glass Health's current comparison material, checked on 15 October 2026, describes documentation, differential diagnosis, planning and clinical questions. Its offer should not be reduced to a single differential-generator feature. Likewise, Doximity's clinician AI is now presented as Doximity Ask, previously DoxGPT; PeerCheck concerns physician review of answers, not a guarantee attached to every future response.
Microsoft's healthcare portfolio also needs product-level description. Dragon Copilot's contextual content offering is relevant to documentation and access to medical information within work. The earlier assertion of a universal six-figure annual minimum for "Copilot for Healthcare" was not supported by a verified public price schedule. Organisations should obtain a scoped quotation rather than infer a single price or deployment route for Microsoft's different healthcare offerings.
The UK gap, challenge and opportunity
The availability question is narrower than the earlier assertion that none of OpenAI's products or Astra could be used by individual UK clinicians. Health-specific eligibility, enterprise procurement, general ChatGPT access, model access and permission to process patient data are different questions. The current Clinicians eligibility documentation is US-focused; it does not establish a blanket prohibition on general ChatGPT or model use in Britain.
The September public-data list does reveal an important localisation issue. PubMed and ClinicalTrials.gov have international relevance, while the remaining named services principally reflect US medicines, payment and professional infrastructure. The release did not announce NICE, CKS, SIGN, UK SmPC information from emc or local NHS pathways as equivalents within that nine-source plugin. It also did not announce an EMIS or SystmOne connection equivalent to the specified enterprise Epic integration. That is a statement about this documented release, not a claim that general AI cannot find UK guidance.
For a fictional UK GP reviewing a proposed diabetes-management explanation, the question is not whether a global model recognises diabetes. It is whether the recommendation fits current NICE guidance, the patient's circumstances and the relevant local pathway. NICE NG28 and US ADA guidance should not be assumed interchangeable. A correctly cited recommendation can still be inappropriate to the setting.
Regulatory assessment follows function and intended purpose. MHRA guidance does not require every educational or administrative application to follow one identical medical-device route. Where applicable, product conformity assessment, registration, information governance and the English NHS's manufacturer and deployment clinical-safety standards answer different questions. Regulatory status should be stated for the relevant entity, function, configuration and jurisdiction, not used as an unexplained brand-wide badge.
The UK's policy position has also moved beyond recommendations alone. In its 6 October 2026 response, the Government accepted the Commission's 44 recommendations, while further implementation work remained. Acceptance is not equivalent to every proposed mechanism already being in force.
iatroX's UK-guidance focus is a relevant product distinction, not proof of superior outcomes or an assumed regulatory designation. The earlier prediction that iatroX would already hold a particular regulatory status when OpenAI expanded was not established evidence and should not be treated as such. As described on 15 October 2026, Ask-iatroX provides free reference access grounded in NICE, CKS, SIGN and SmPC material from emc, with linked sources and a published methodology. Its design supports inspection; clinicians must still assess the answer.
What this means for the market
Free professional access could pressure the price of basic clinical-answer generation, but that is a strategic hypothesis rather than evidence that every specialist subscription must disappear. Organisations still incur integration, assurance, support and review costs. Clinicians may also value a specialised learning sequence independently of whether a general assistant can answer the same question.
The distinction must be fair to general-purpose products. ChatGPT supports Study Mode's guided questioning, and can assist with calculation explanations, examination study and reflective drafting. The earlier claim that it simply does not offer learning or CPD-related functionality was too broad, particularly given ChatGPT for Clinicians' published CME offering. Actual calculations, credit eligibility and professional records require their own checks.
iatroX's more specific proposition is the connection between an attempted medical question, targeted Socratic follow-up, adaptive and spaced practice, exam-date-aware planning, simulations and reviewed learning records. These are the platform's published design features as of 15 October 2026, not evidence that it produces better educational outcomes than ChatGPT or the Johns Hopkins collaboration. No iatroX connection to Epic, OpenEvidence or OpenAI's healthcare products is implied.
The scenario-based verdict is therefore straightforward. An eligible US clinician needing a broad professional assistant can assess ChatGPT for Clinicians against that task. A hospital needing authorised patient-context access must assess the institutional product and its actual integration. A patient exploring their own connected information is considering Health, with different controls and limits. A clinician seeking UK-guidance reference and structured professional learning may find iatroX relevant without needing to replace their record system. The right comparison is the completed task and the work left for the user, not the breadth of the company name.
The dated timeline
| Date | Published development | Boundary to retain |
|---|---|---|
| 7 January 2026 | ChatGPT Health introduced with limited access | Consumer support, not unrestricted clinical practice |
| 8 January 2026 | ChatGPT for Healthcare announced | Organisational deployment and controls |
| 21 January 2026 | OpenEvidence announced $250 million Series D at $12 billion | Historical financing, not September's valuation |
| 22 April 2026 | ChatGPT for Clinicians announced | Free access for eligible verified US professionals |
| 23 July 2026 | Broader Health rollout announced | US adults on supported web and iOS plans, with permissioned context |
| 1 September 2026 | Enterprise Epic and public-data connections announced | Read-only Epic pilot versus separately available public sources |
| 3 September 2026 | GPT-6 Astra announced | A model release does not identify every healthcare deployment's model |
| September 2026 | OpenEvidence's further $250 million financing reported at $15 billion | Reported transaction and valuation, not clinical validation |
| 29 September 2026 | GPT-6.1 Sol documentation published | A later model milestone than Astra |
| 13 October 2026 | OpenEvidence and Johns Hopkins Medicine announced collaboration | Clinical and educational ambition, not completed outcome evidence |
The linked announcements above support the timeline. Keeping the dates visible allows later product changes to be assessed against the right starting point instead of silently applying a new capability to an older evaluation.
Turn this reading into evidence you can use
A useful learning activity is to compare one vendor's product claim with its supporting evaluation, identify the actual task and denominator, then explain what remains uncertain. Use fictional or properly de-identified material, and do not transfer patient records into an unrelated learning tool.
As described on 15 October 2026, iatroX CPD connects learning activities with a draft record that the clinician reviews and personalises, including suggested framework mapping where relevant. CPD tools are included in the paid learning bundle rather than a separate add-on; completed evidence remains accessible after the subscription ends. A personally reviewed learning record is not automatically accredited CME, and it should describe what the clinician genuinely studied and learned.
Frequently asked questions
Does ChatGPT Health have the same Epic connection as ChatGPT for Healthcare?
No: patient-connected Health information and the enterprise Epic plugin use different access arrangements. The September 2026 institutional connection is read-only and subject to the organisation's configuration and existing permissions.
Does GPT-6 Astra power every OpenAI healthcare product?
That cannot be inferred from Astra's model announcement, and a later GPT-6.1 Sol release does not establish it either. The model and configuration used in a particular deployment must be checked in its current product documentation.
Has independent research established the safety of every OpenAI healthcare workflow?
No single study establishes that: the February 2026 ChatGPT Health triage experiment and September's company-reported EHR response evaluation concern different tasks and endpoints. Neither should be treated as blanket validation or invalidation of every healthcare use.