Table of Contents
OpenAI has expanded ChatGPT for Healthcare with computer-use agents that navigate electronic health records and insurer portals, improved clinical search, and more than 20 new healthcare plugins. The agents can read screens, click, type, and carry information between applications, including software whose APIs do not support all the steps a care team needs to complete.
The healthcare update connects medical evidence and patient information with tools for referrals, documentation, insurance checks, and care coordination. CVS Health and Cigna Healthcare are already using computer-use capabilities, while Abridge is developing agents for discharge coordination and Ambience has demonstrated approval controls for prior authorizations.
OpenAI also published new evaluations of GPT-6 Astra and GPT-6 Luna on clinical and administrative work. Its results include physician reviews of EHR inbox tasks and a cost comparison on a Stanford-developed healthcare administration benchmark.
Moving From a Patient Message to an Authorization Request
Computer use addresses a practical integration problem. A patient’s records may contain the information needed for care, but staff still have to retrieve it, open another application, and enter it again. APIs, the software interfaces that let systems exchange information directly, can automate some of that work. Screen-based agents can handle steps available through an application’s interface when those integrations fall short.
In OpenAI’s demonstration using synthetic patient data, a clinician uses Voice in the ChatGPT app to ask for help with a back-pain MRI request. The agent reviews visit notes, physical therapy records, and previous imaging in the EHR, then drafts an MRI order, an encounter note, and a patient reply for the clinician to review.
The workflow continues into the insurer’s portal, where the agent uses the same records to prepare a prior authorization request for an authorization specialist’s review. Prior authorization is the process of obtaining an insurer’s approval before a service is covered; completing it can require assembling clinical evidence and entering it into a separate payer system.
The new plugins provide another route to working across healthcare tools within a ChatGPT conversation. They include UpToDate Expert AI, Luma Health, Stedi, Redox, Elation Health, Oracle HDI, Turquoise, and IMO Health, spanning clinical information, patient coordination, insurance, and operations. The plugins are also available through ChartGPT in the EHR workflow.
One example combines Luma Health’s appointment information with TriZetto eligibility checks to prepare a call list of patients who need confirmation or coverage issues resolved. Another uses IMO Health to identify an MRI procedure code, Stedi to check benefits, and Turquoise to compare negotiated rates at two hospitals.
Trusted clinical search has also been updated in ChatGPT for Healthcare and ChatGPT for Clinicians to improve how it finds, evaluates, and cites medical evidence. In a blind evaluation involving more than 1,300 physician reviews across multiple specialties, OpenAI said the updated experience received higher ratings across all eight measured dimensions, including factual correctness, safety, actionability, and source authoritativeness.
What the Clinical and Administrative Evaluations Measured
OpenAI’s computer-use evaluations tested GPT-6 Astra on EHR inbox tasks shaped by physician input. The work required reasoning over a patient’s history while navigating interfaces and producing notes, orders, and messages.
Reviewers assessed clinical safety by examining whether inaccuracies, omissions, or actions could plausibly cause patient harm. They rated 99% of 325 reviewed tasks safe. In a separate helpfulness assessment, physicians selected “very helpful” or “extremely helpful” in 93% of 646 assessments.
For administrative work, OpenAI tested the models on HealthAdminBench, an independent benchmark developed by Stanford researchers and collaborators. It includes 135 tasks covering prior authorizations, durable medical equipment orders, and appeals, with agents working through browser-based healthcare environments.
| Model | Available Rubric Points Earned | Reported Cost per Task |
|---|---|---|
| GPT-6 Astra | 94.3% | — |
| GPT-6 Luna | 91.8% | $0.085 |
Both models ran at maximum reasoning effort in OpenAI’s evaluation. Luna’s reported cost was 98% lower than Astra’s, while its rubric score was 2.5 percentage points lower. These figures measure the share of available rubric points earned, rather than the percentage of complete workflows finished successfully.
How Care Organizations Are Putting Agents to Work
CVS Health uses computer-use capabilities to send documents, process referrals, and support specialist consultations across existing clinical applications. OpenAI reported that one production workflow cleared a months-long document-sending backlog over a weekend, while another handles system steps for tens of thousands of consultations each day. CVS uses validation checks, human oversight, screenshots, and action records to verify the work.
CVS has also released its Apache 2.0-licensed CUA Agent Kit. Built with OpenAI’s Agents SDK, the kit uses a computer-use agent to discover a workflow from plain-language instructions and record successful runs. A Codex skill then turns that evidence into repeatable Playwright browser automation or pywinauto Windows automation that can run without further model calls.
That separates discovering a process from repeatedly executing it. Teams can inspect and maintain the generated code, and use production traces to help Codex diagnose broken workflows and propose updates for engineering review.
At Cigna Healthcare, care managers previously navigated more than 10 clinical and administrative systems to gather information about benefits, providers, appointments, and prior interactions. Computer use now helps bring relevant information together in seconds, OpenAI reported. Cigna Healthcare CIO Madhu Nutakki said the project moved from concept to deployment in under eight weeks.
Abridge’s planned inpatient application focuses on barriers to discharge. Agents would combine clinical conversations, EHR information, and staff context to coordinate medication access and home-health services, using computer interaction alongside calls and messages. Clinicians would remain involved in decisions requiring clinical judgment.
Ambience’s prior-authorization demonstration focuses on controlling actions as well as completing them. When the patient record is ambiguous, its agent stops and asks the clinician, then waits for approval before submitting. Health systems can determine permitted actions, pause the agent, and inspect a step-by-step audit log.
Access Controls and Availability
ChatGPT for Healthcare provides role-based access controls and audit logs, with organizations able to configure agent access and add review and approval steps. Organization data is not used to train OpenAI foundation models.
Workflows involving protected health information require eligible services covered by a Business Associate Agreement with OpenAI, the required security and retention settings, and protections extending to connected systems. Access to an agent does not replace those deployment requirements.
Existing ChatGPT for Healthcare customers can access computer use in ChatGPT Work and Codex, along with improved clinical search, subject to workspace settings. Administrators can enable the new healthcare plugins. Eligible, verified U.S. clinicians can use a free ChatGPT for Clinicians account, which includes computer use, improved search, and select plugins. Developers can build computer-use applications through the OpenAI API.