On November 16, 2026, OpenAI announced at the HLTH26 conference that it is bringing computer-use agents into healthcare: the agents read screens, click, and type inside the electronic health record and insurance systems care teams already use, taking over cross-system paperwork such as referrals, prior authorizations, discharge planning, and follow-up outreach. The same day, ChatGPT for Healthcare gained more than 20 new plugins, with clinical search upgraded in parallel.
No New System — the Agent Works Inside the Old Software
The idea is not to make hospitals switch systems, but to teach agents to use the software they already have: reading screens, clicking buttons, filling in forms, and moving information between applications. There are three entry points: the OpenAI API, ChatGPT for Healthcare, and the free ChatGPT for Clinicians; existing customers enable it in ChatGPT Work and Codex through a workspace administrator. The first partners are Abridge, Ambience, Cigna Healthcare, and CVS Health, covering referrals, authorizations, and document handling. CVS Health is furthest along: one production agent cleared a months-long document-sending backlog over a single weekend, another handles the system steps behind tens of thousands of specialist consultations a day, and CVS has open-sourced its self-built Agent Kit.
99% Safe — But That Is a Physician Review, Not a Patient Outcome
OpenAI also published evaluation figures for GPT-6 Astra: in 325 electronic health record inbox tasks, 99% were judged clinically safe by physician reviewers, and in 646 assessments, 93% of the work was rated 'very helpful' or 'extremely helpful.' On HealthAdminBench, an independent benchmark from Stanford researchers, Astra scored 94.3% of the available points, while the cheaper GPT-6 Luna scored 91.8% at 8.5 cents per task — 98% cheaper than Astra. These numbers need to be read correctly: 99% is physicians' assessment of the model's work, not measured patient outcomes. For the remaining 1%, OpenAI's health lead said researchers found no case where the model created a specific, plausible risk of patient harm. The compliance boundaries are also clearly drawn: content is not used to train models, use involving protected health information requires a business associate agreement and HIPAA compliance, and plugins must be enabled one by one by an administrator.
Plugins are the other half of this upgrade: more than 20 new plugins cover scheduling, referrals, records, insurance verification, and cost analysis, including UpToDate Expert AI, Availity Extend, and Oracle HDI, while ChartGPT can bring ChatGPT into the electronic health record workflow. For hospitals, the real test starts at go-live: every step an agent takes in production must be logged and auditable, and who is responsible when something goes wrong — and whether review processes keep up — matters far more than how smooth the demo looked. What is worth watching next is production data and audit results from more hospitals, not the numbers from the launch event.