OpenAI announced on July 23, 2026 that Health in ChatGPT is rolling out to logged-in U.S. users aged 18 and older on the web and iOS. Users can choose to connect Apple Health and supported medical records so ChatGPT can use personal context when permission is granted. The feature is not yet available in Codex, and it should not be treated as a diagnostic system.
The product addresses a practical fragmentation problem. Health information can sit across patient portals, medical records, fitness apps, wearables, and uploaded files. OpenAI says that, once connected, ChatGPT can help compare new and previous test results, summarize changes before an appointment, or consider sleep, activity, and workouts together. Those are provider descriptions of intended use; the data available in practice depends on the connected service, region, and permissions.
Permission is central to the design. OpenAI says ChatGPT asks before using connected medical records and Apple Health information by default. Users can approve individual requests, allow ongoing access, or disconnect an account at any time. After disconnection, data synced into OpenAI’s systems is deleted within 30 days, while information already included in ChatGPT conversation history must be deleted separately. That distinction matters before a family or organization treats the feature as a records system.
OpenAI also says connected medical records, Apple Health information, and conversations that use them are not used to train foundation models or target ads. Health adds encryption protections and safeguards for sensitive sharing. These are the provider’s current privacy and product commitments, so users should still review the applicable privacy notice, connected-service terms, retention behavior, and regional availability before relying on them.
On model performance, OpenAI says GPT-5.5 Instant and GPT-5.6 Sol improve health conversations, with HealthBench Professional and physician-developed scenarios among the evaluation inputs. The evaluations and early rollout are useful signals, but they do not establish reliability for every symptom, medication, or test result. OpenAI explicitly says ChatGPT can make mistakes and does not replace qualified medical care or clinical judgment.
The broader product shift is from one-off health questions to a persistent personal-context workflow. With permission, information can be referenced, corrected, and managed across conversations, while users can ask the system to prepare questions for a clinician. The hard engineering and governance problems are data freshness, source completeness, granular access, auditability, and a clear escalation path when a question requires professional care.
Health in ChatGPT shows personal AI moving from answering questions to maintaining context across data sources. That can reduce repetitive preparation, but a richer context also creates a larger data-governance obligation. Adoption should begin with explicit consent, revocable access, source checking, and clinical oversight rather than assuming that a polished demo proves safety.
