News

OpenAI makes computer-use agents available to healthcare teams

OpenAI has made computer-use capabilities available through its API and healthcare products, while safety, cost, performance and customer-impact figures remain company claims.

D
Nov 16, 2026 · 3 min read

OpenAI made computer-use capabilities available to healthcare teams, developers and individual clinicians on Monday through its API and healthcare products. The release lets users deploy agents that navigate existing browser and desktop interfaces instead of rebuilding each workflow around a new integration.

ChatGPT for Healthcare customers and individual clinicians can use the agents in ChatGPT Work and Codex, OpenAI said. Customer access depends on workspace settings, while developers can use the capability through the API.

The agents read screens, click and type, then return screenshots. OpenAI’s computer-use guidance says the integrating application remains responsible for the computer session, permissions, execution of actions and verification of results. The company recommends isolating the environment, limiting allowed sites and actions, requiring confirmation for consequential steps, bounding each run and checking the outcome.

OpenAI said ChatGPT for Healthcare includes role-based access controls and audit logs. It also said workflows involving protected health information must use eligible services covered by a Business Associate Agreement and the required security and data-retention settings.

The release also adds more than 20 healthcare plugins, extending the workflow-integration approach seen in OpenAI’s expanded ServiceNow partnership. OpenAI said all of the new healthcare plugins were available in ChatGPT for Healthcare at launch, while only a selection was available to individual ChatGPT for Clinicians users. The company also announced changes to clinical search in both products.

OpenAI said physicians preferred the updated clinical-search experience across all eight measured dimensions in more than 1,300 reviews. Separately, the company said 99% of 325 reviewed electronic-health-record inbox tasks were rated clinically safe and that physicians rated the output very or extremely helpful in 93% of 646 assessments. Those results have not been independently validated, and the opened materials did not provide enough detail to assess the underlying data, review protocol, error distribution or patient-harm methodology.

For administrative work, OpenAI said GPT-6 Astra earned 94.3% and GPT-6 Luna earned 91.8% of available rubric points on HealthAdminBench. It also said Luna cost 8.5 cents per task, 98% less, but the benchmark materials did not independently reproduce the new model results or cost calculation.

HealthAdminBench contains 135 expert-defined tasks in four simulated graphical environments, covering prior authorization, appeals and denials, and durable-medical-equipment order processing. The April paper reported a best end-to-end task-success rate of 36.3% and a highest subtask-success rate of 82.8% among the configurations it evaluated. Those measures differ from OpenAI’s percentage of available rubric points, so the figures are not directly comparable.

OpenAI named Abridge, Ambience, Cigna Healthcare and CVS Health as organizations it is working with on healthcare computer-use applications. It said one CVS agent cleared a months-long document backlog over a weekend and another handles system steps for tens of thousands of consults each day. Those impact figures were not independently established by the opened CVS materials.

CVS Health has published an Apache-2.0-licensed CUA Agent Kit that records successful computer-use discovery runs and converts them into deterministic Playwright or pywinauto code. The generated workflow is designed to run without repeated model calls, but the repository corroborates workflow categories rather than OpenAI’s production-scale claims.

The materials reviewed did not specify complete pricing, organizational eligibility, geographic availability, rollout limits or model-level access requirements for every product surface.

More news