Summary:
The expansion connects more clinical and administrative tools. Access settings, approval points and separate evaluation denominators matter when planning a pilot.OpenAI’s November 16 update expands ChatGPT for Healthcare with computer-use agents, improved clinical search and more than 20 new healthcare plugins. The practical focus is work spread across records, scheduling and administrative systems, with organizations deciding which applications agents may access and where human review is required.
What is available
OpenAI says the agents and search improvements are available in ChatGPT Work and Codex, subject to workspace settings. Healthcare customers can ask administrators to enable the new plugins. Eligible, verified U.S. clinicians can use the free ChatGPT for Clinicians product, where selected plugins are available. Healthcare organizations can contact sales, including for an AWS Marketplace private offer.
The plugin lineup includes Amion, Metriport, Stedi and UpToDate Expert AI. Examples span finding records and policies, coordinating outreach and checking eligibility. Deployment involving protected health information requires eligible services covered by a Business Associate Agreement, appropriate security and retention settings, and protections for connected systems. Source: OpenAI’s product announcement.
Read the evaluation numbers with their denominators
In its separate computer-use report, OpenAI says physicians rated 99% of 325 reviewed tasks safe. It also reports that 93% of 646 helpfulness assessments received one of the top two ratings. These are different evaluation denominators, not a single clinical success rate.
On Stanford’s HealthAdminBench, OpenAI reports GPT-6 Astra earned 94.3% of available rubric points and GPT-6 Luna 91.8%, both at maximum reasoning effort. Luna’s reported cost was 8.5 cents per task, 98% below Astra’s. The benchmark is independently developed; these model results are reported by OpenAI. They do not measure patient outcomes or guarantee performance in a hospital’s own systems.
The report also distinguishes deployments from development work. CVS Health describes production administrative workflows with validation and human oversight. Abridge’s inpatient example is planned work. Keeping those stages separate matters when assessing how readily a demonstration can be adopted.
Start with an auditable handoff
For a pilot, a useful first question is where the agent’s work ends and a person’s decision begins. Choose one bounded workflow, define the permitted systems and actions, and make the approval point explicit. Record both successful completions and interventions: a fast draft that repeatedly needs correction may save less time than its demonstration suggests.
Teams comparing approaches can also read our coverage of Gemini for healthcare, while checking each product’s current access and controls separately.
Sources and image
- OpenAI: Healthcare product update, November 16, 2026
- OpenAI: Computer-use evaluations and customer examples, November 16, 2026
Cover: original conceptual illustration of a healthcare information handoff and human review, not an actual product interface or clinical event.


