OpenAI Says GPT-6 Astra Cleared 99% of Its Own Medical Safety Tests
OpenAI reviewed 325 electronic health-record inbox tasks and rated GPT-6 Astra clinically safe on 99% of them — by its own reckoning. The Nov. 16 disclosure landed alongside a computer-use update that lets AI agents click, type and hop between medical apps. OpenAI's Karan Singhal told Fierce Healthcare the model can still overreach and act without proper sign-off.
- Original source: thedefiant.io.
- The source published it on Nov 17, 2026 at 12:20 AM.
- No second outlet has reported it yet. Treat it as a single-source claim.
OpenAI put its GPT-6 Astra agent through 325 reviewed medical-inbox tasks and gave itself a 99% clinically safe rating on them. The evaluation, published Nov. 16, covered agents working inside electronic health-record software. Reviewers looked at what the model did alongside the notes, orders and patient messages it produced, then asked whether inaccuracies, omissions or actions could plausibly hurt a patient.
On the cases that didn't earn the safe label, OpenAI's Karan Singhal told Fierce Healthcare the researchers hadn't found any instance where the model created a specific, plausible risk of patient harm. His caveat was blunter: agents can wander out of bounds. “In some cases, the model might overreach and might take an action without getting the proper sign-off or authorization,” he said.
That 99% number is a company-reported assessment of task work — not a measured patient-outcome rate, and not independent clinical validation. Physicians helped shape the tasks, but OpenAI's announcement never names the safety reviewers or says whether they stood outside the company. It also leaves unanswered how the 325 cases were picked, whether the patient data was real or synthetic, and whether every order and message stayed a draft for a clinician to approve.
The safety writeup arrived with OpenAI's Nov. 16 computer-use update for healthcare, which pushes agents beyond reading and into doing — clicking, typing and moving between applications to complete multi-step work. It builds on the Sept. 1 Epic integration that piped authorized patient-record information into ChatGPT so clinicians could review histories, spot changes and prep for appointments.
In a separate demo using synthetic patient data, OpenAI showed an agent handling a lower back pain report: it reviewed records, drafted an MRI order, an encounter note and a patient reply for clinician review, then queued up an insurance authorization request for a specialist to approve. That's a supervised workflow by design — and OpenAI says it doesn't show how approvals actually worked across the safety-evaluation sample.
The healthcare computer-use features are available through OpenAI's API, ChatGPT for Healthcare and ChatGPT for Clinicians. Developer guidance instructs builders to ask the user right before taking medical-care actions, or before sending, posting or submitting information to a third party on the user's behalf.
TopicsOpenAI · GPT-6 Astra · Karan Singhal · Fierce Healthcare · Epic · ChatGPT for Healthcare · ChatGPT for Clinicians
Source: thedefiant.io
Written by AI from verified reporting.
More stories
- Tech & AI blockchain.news 0 viewsGrid Robotics paper pitches 'agentic robotics' powered by GPT-6 Astra
- Tech & AI x.com 0 viewsNvidia director warns board of losing "irreplaceable talent" if Huang isn't reinstated as CEO
- Tech & AI blockchain.news 0 viewsNVIDIA's Vera Rubin aces MLPerf debut with 10x agentic efficiency gain
- Tech & AI blockchain.news 0 viewsOpenAI Adds Computer-Use Agents and 20+ Plugins to ChatGPT for Healthcare

