OpenAI has brought computer-use agents to healthcare, letting its models operate the electronic health record systems, insurer portals and administrative software that hospitals and payers already run by clicking and typing through them. CVS Health, Cigna Healthcare, Abridge and Ambience Healthcare are the first named partners. In OpenAI's own reviews, GPT-6 Astra was rated clinically safe on 99% of 325 EHR inbox tasks.
The agents are available in ChatGPT for Healthcare, in ChatGPT for Clinicians and through the OpenAI API. OpenAI paired the launch on Nov. 16 with more than 20 new healthcare plugins and improved clinical answers. ChatGPT for Clinicians remains free for verified US clinicians. OpenAI did not publish a price for the healthcare workspace.
Because the agents work through existing screens rather than dedicated integrations, they can handle multistep jobs where APIs do not fully cover a function, such as authorization requests, referrals and discharge planning. In a demonstration using a synthetic patient, a message about lower back pain led the agent to review visit notes, physical therapy records and prior imaging. It then drafted an MRI order, an encounter note and a reply to the patient for clinician review, and prepared a prior authorization request in an insurer portal for an authorization specialist to check.
Key facts
- 01OpenAI launched healthcare computer-use agents on Nov. 16 in ChatGPT for Healthcare, ChatGPT for Clinicians and its API.
- 02Reviewers rated GPT-6 Astra clinically safe on 99% of 325 EHR inbox tasks; physicians rated it highly helpful in 93% of 646 assessments.
- 03On Stanford's HealthAdminBench, GPT-6 Astra earned 94.3% of rubric points and GPT-6 Luna 91.8%, at 8.5 cents per task for Luna.
- 04Early partners are CVS Health, Cigna Healthcare, Abridge and Ambience Healthcare; OpenAI also added 20+ healthcare plugins.
- 05CVS Health open-sourced a CUA Agent Kit under Apache-2.0 that turns successful agent runs into scripts that run without model calls.
OpenAI's safety review asked whether any inaccuracy, omission or action could plausibly lead to patient harm, and reviewers examined the agent's interface actions alongside the notes, orders and messages it produced. On helpfulness, physicians placed GPT-6 Astra in the top two categories of a five-point scale, very helpful or extremely helpful, in 93% of 646 assessments.
OpenAI also reported results on HealthAdminBench, a Stanford-built benchmark of 135 tasks and 1,698 subtasks covering prior authorization, durable medical equipment orders and appeals. GPT-6 Astra earned 94.3% of available rubric points and the smaller GPT-6 Luna earned 91.8%, with both running at maximum reasoning effort. Luna completed each task for about 8.5 cents, 98% less than Astra.
CVS Health is already running agents in production to send documents and process referrals. One agent cleared a document-sending backlog that had built up over months in a single weekend, and another handles the system steps for tens of thousands of specialist consults each day. CVS Health also published a CUA Agent Kit under the Apache-2.0 license that uses an OpenAI computer-use agent to discover browser or Windows workflows, then converts successful runs into Playwright or pywinauto code that executes without calling a model.
At Cigna Healthcare, case managers previously moved between more than 10 separate clinical and administrative systems before contacting a member. The new deployment assembles coverage, provider, appointment and prior-interaction details in one place, and Cigna took it from concept to deployment in under eight weeks as part of The Cigna Group's broader collaboration with OpenAI.
Abridge is building a cross-system assistant for discharge work spanning medication access and home health, running concurrent agents that escalate to clinicians when judgment is needed. Ambience Healthcare's agent prepares prior authorization requests from chart evidence, stops to ask a clinician when the record is ambiguous and waits for approval before submitting. Health systems using Ambience can restrict what the agent is allowed to do, pause it and inspect a step-by-step audit log.
The benchmark numbers come with limits. The HealthAdminBench scores OpenAI cited are rubric credit, which awards partial points, while the benchmark's original paper found the best end-to-end task success among seven tested configurations was 36.3%. That paper also simplified logins, CAPTCHAs and session timeouts, and OpenAI did not disclose an end-to-end success rate for the new models.
ChatGPT for Healthcare includes role-based access controls and audit logs, and organizations can add approval steps and limit agents to the apps, records and actions each task requires. HIPAA workflows require eligible services covered by a business associate agreement with OpenAI, workspace admins must switch on plugins such as Availity, Oracle HDI and UpToDate Expert AI, and OpenAI says organization data is not used to train its foundation models.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.




