OpenAI says GPT-6 Astra, its computer-use model, was rated safe on 99% of 325 reviewed clinical inbox tasks, and that a cheaper sibling called GPT-6 Luna scores within three points of it on an administrative benchmark for 8.5 cents a task. The figures come from OpenAI’s own post on computer-use agents for healthcare teams, published on openai.com.

The problem it targets is mundane. A clinician recommends an MRI, the insurer needs facts that sit in the patient’s record, and a staff member copies them by hand into a separate portal. A computer-use agent does that work through the screen, with mouse clicks and keystrokes across applications, so it can operate software that offers no API.

OpenAI built a demo with fake patient data. A patient writes in about lower back pain. The agent reads the EHR, covering the visit notes, physical therapy history, and earlier scans. It drafts an MRI order, a note for the encounter, and a reply to the patient, all for the clinician to review. It then assembles the prior authorization in the insurer’s portal for a specialist to check. Nothing goes out unreviewed.

On the clinical side, OpenAI had reviewers examine tasks that clinicians handle in their EHR inboxes, shaped by physician input. Safety meant asking whether an error, omission, or action might realistically harm a patient. Physicians called the work “very helpful” or “extremely helpful” in 93% of 646 assessments.

The administrative numbers use an outside yardstick. HealthAdminBench comes from Stanford researchers and is independent of OpenAI. It covers prior authorizations, appeals, and equipment orders. Astra earned 94.3% of the rubric points on offer. Luna earned 91.8% while costing 98% less, which works out to 8.5 cents for each task. Maximum reasoning effort was used for both.

OpenAI named four partners, and the descriptions are theirs, not independent audits.

The controls matter as much as the scores. ChatGPT for Healthcare offers role-based access and audit logs, and OpenAI says organization data does not train its foundation models. HIPAA-compliant workflows depend on eligible services under a Business Associate Agreement with OpenAI, plus the security and retention configuration that agreement demands.

Availability is broad. The capabilities are live in the OpenAI API, ChatGPT for Healthcare, and ChatGPT for Clinicians, and a free ChatGPT for Clinicians account is open to U.S. clinicians.

Two gaps stand out. The 99% safety rating comes from reviewers OpenAI selected, on tasks it designed, and the post does not name who the reviewers were. The CVS Health and Cigna results are also reported by OpenAI, with no baseline for error rates before the agents arrived.

Health systems weighing a pilot should ask for their own error rates on a month of real authorization requests before they accept either figure, and should treat the 8.5 cents per task as the price to beat when negotiating.

OpenAI, “Advancing computer-use agents for healthcare teams,” openai.com, publication date not shown on the page.