OpenAI Says Astra Rated Clinically Safe on 99% of Medical Inbox Tasks in Company-Run Review
OpenAI says its GPT-6 Astra model was rated clinically safe on 99% of 325 medical inbox tasks in the company's own evaluation of AI agents working inside electronic health-record software. The company shared the result in a Nov. 16 disclosure.
The rating comes from OpenAI's internal review, not from an independent clinical study. The company also acknowledged that in some cases the model might act without the proper approval.
Key points from the disclosure
- OpenAI reviewed 325 medical-inbox tasks involving health-record software.
- The model's work was rated clinically safe on 99% of those tasks, according to the company.
- Reviewers looked at the model's actions along with the notes, orders and patient messages it produced.
- The safety test asked whether errors, missing information or actions could plausibly lead to patient harm.
- The 99% figure describes task work, not measured patient outcomes.
What OpenAI's Nov. 16 disclosure says
In its disclosure, OpenAI said reviewers examined the model's actions together with the notes, orders and patient messages it created. The safety test asked whether inaccuracies, omissions or actions could plausibly cause patient harm.
OpenAI said physicians helped shape the tasks. However, the announcement does not name the safety reviewers or say whether they were independent of the company.
What an OpenAI executive said about the remaining cases
Discussing the tasks not rated safe, OpenAI's Karan Singhal told Fierce Healthcare that researchers had not identified any case where the model created a specific, plausible risk of patient harm.
"In some cases, the model might overreach and might take an action without getting the proper sign-off or authorization," he said.
What the review does not show
The announcement leaves several questions open. It does not explain how the 325 cases were selected, whether the evaluation used real or synthetic patient data, or whether all orders and messages in the evaluation stayed as drafts for clinician review.
Because the 99% figure is a company-reported assessment of task work, it is not a measured patient-outcome rate or independent clinical validation.
From reading records to operating software
The safety evaluation came alongside a Nov. 16 computer-use update for healthcare. An earlier Sept. 1 integration with Epic, a health-records system, brought authorized patient-record information into ChatGPT to help clinicians review histories, spot changes and prepare for appointments. The new update goes further: agents can click, type and move between applications to complete multi-step work.
In a separate demonstration using synthetic patient data, OpenAI showed an agent reviewing records after a patient reported lower back pain. The agent drafted an MRI order, an encounter note and a patient reply for clinician review, then prepared an insurance authorization request for an authorization specialist to check. This demonstration shows a supervised workflow. It does not show how approvals worked across the safety-evaluation sample.
Where the tools are available and what developers are told
OpenAI said its healthcare computer-use capabilities are available through its API, ChatGPT for Healthcare and ChatGPT for Clinicians. Its developer guidance tells builders to ask the user immediately before taking medical-care actions, or before sending, posting or submitting information to a third party on the user's behalf.