OpenAI is rolling out its dedicated ChatGPT health experience to adults in the United States, allowing users to connect medical records, Apple Health data and supported wellness apps. The system can help interpret lab results, organize questions for a medical appointment and review information such as sleep and activity patterns, but the underlying model differs depending on whether a user pays for ChatGPT.
Free users receive health responses generated by GPT-5.5 Instant, while subscribers can access GPT-5.6 Sol, OpenAI’s newer flagship model. GPT-5.6 Sol performs better on the company’s health evaluations, creating a consequential capability gap in a setting where incomplete, misleading or overly confident answers can affect real-world decisions.
OpenAI says both models surpass physicians’ answers on HealthBench Professional, a benchmark designed to evaluate AI systems on realistic medical tasks. Such results do not establish that a chatbot can replace a clinician, however. Controlled tests primarily measure how well a system handles predefined questions and scoring criteria. They cannot reproduce a physical examination, longitudinal knowledge of a patient, access to a medical team or the nonverbal signals that can shape a diagnosis.
The company repeatedly warns that ChatGPT may make mistakes and is not a substitute for professional medical advice. More than 260 physicians contributed to the development of its health features. OpenAI also says connected health information will not be used to train its models or for advertising, an important distinction given the sensitivity of medical records and data collected by consumer devices.
The health tools were initially built around a separate area within ChatGPT, where users could manage connected information and revisit earlier health conversations. Early testing found that more than 70% of participants continued asking medical questions in ordinary chats because moving into the dedicated section added friction. OpenAI has therefore made the health experience available from any conversation while retaining the separate space for data management and health-chat history.
The rollout follows rapid growth in medical use of the chatbot. OpenAI says more than 300 million people now ask ChatGPT health-related questions each week, compared with 230 million when the health product was introduced in January. That scale raises the stakes of model selection, safety warnings and the presentation of uncertainty, particularly when stronger performance is tied to a subscription.
Medical AI benchmarks have produced mixed results across different tasks. Systems including MIRA, which works with electronic health records, and AMIE have performed roughly on par with primary care physicians in simulated consultations. Yet a more recent radiology evaluation, RadLE 2.0, found that none of the 16 tested AI models matched human radiologists. A central weakness was that chatbots could deliver incorrect findings confidently, while specialists were more likely to recognize and communicate uncertainty.
That limitation is especially important in consumer chat systems, whose fluent language can make an answer sound more authoritative than the available evidence supports. A chatbot also lacks the ability to directly examine a patient, order appropriate tests on its own or reliably recognize every emergency from a short written description. Health tools can still assist with routine preparation and information review, but responsibility for diagnosis and treatment remains with qualified medical professionals.
OpenAI has not announced whether or when the product will reach Europe. Its January unveiling excluded the European Economic Area, Switzerland and the United Kingdom, where health-data protections and rules governing higher-risk AI applications could complicate deployment. For now, access is expanding to U.S. users aged 18 and older, with the quality of the model available for health conversations determined in part by subscription status.
Sources: The Decoder