Products#AI safety#Device control

OpenAI brings computer-use agents to hospitals and prior auth

OpenAI added computer-use agents to ChatGPT for Healthcare on Nov 16: 99% of 325 clinical tasks rated safe, 8.5 cents per admin task, and 21 plugins.

A clinician reviews medical images and records at a two-monitor workstation

OpenAI put computer-use agents into ChatGPT for Healthcare on November 16. Instead of waiting for hospitals to expose every step as an API, the agents read screens, click, and type across electronic health records and payer portals to complete prior authorizations, referrals, and discharge planning. OpenAI also published evaluation numbers: 99% of 325 reviewed clinical tasks were rated safe, and administrative tasks cost 8.5 cents each on HealthAdminBench.

Key points

  • 99% is a task-level safety rating, not a patient outcome. Reviewers examined the model’s actions alongside the notes, orders, and patient messages it produced, judging whether inaccuracies, omissions, or actions could plausibly lead to patient harm. Of 325 reviewed tasks, 99% were rated safe. Physicians separately rated helpfulness “very” or “extremely helpful” in 93% of 646 assessments.
  • 8.5 cents is GPT-6 Luna’s cost per task on HealthAdminBench. The benchmark, maintained by Stanford researchers, covers prior authorizations, durable medical equipment orders, and appeals. Luna earned 91.8% of available rubric points and GPT-6 Astra 94.3%, both at maximum reasoning effort; OpenAI says Luna costs 98% less.
  • The capability is available now. Computer-use agents work today through the OpenAI API, ChatGPT for Healthcare, and ChatGPT for Clinicians. ChatGPT for Healthcare also gains more than 20 healthcare plugins (the list names 21), and CVS Health and Cigna Healthcare are already using the capability in production or pilot workflows.

Background

Healthcare’s payment and delivery systems have never shared one interface. The information an insurer needs for a prior authorization usually sits in the EHR, but a staff member still has to retrieve it and retype it into a separate payer portal. OpenAI had already entered clinical workflows through ChatGPT for Healthcare and ChartGPT; this release adds a route that does not depend on the other side exposing an API, because the agent operates the existing interface.

We covered that trade-off in agent browser control: skipping integration lowers the barrier, but it makes reliability depend on interface stability and permission boundaries. The other route is data plumbing. FHIR MCP Server exposes any FHIR server as tools an agent can read and write, while HealthChain validates FHIR resources against value sets as they are generated. This launch puts both routes side by side: the 21 plugins are integration points, and the computer-use agents work the screen. The first requires an API; the second requires only a UI a person could use.

The facts

OpenAI’s evaluation has two halves. The clinical half used GPT-6 Astra on EHR inbox tasks shaped by physician input, requiring reasoning over longitudinal records and multi-step navigation of complex interfaces. Reviewers looked at the model’s actions together with the notes, orders, and patient messages it produced. Safety was defined as whether inaccuracies, omissions, or actions in the model’s work could plausibly lead to patient harm. That is a task-level judgment; it does not track real patient outcomes. Physicians rated helpfulness on a five-point scale, with 93% of 646 assessments in the top two categories.

The administrative half used HealthAdminBench. According to the benchmark site, it contains 135 tasks, 1,698 subtasks, and four GUI environments drawn from prior authorization, durable medical equipment orders, and appeals, across three difficulty levels. Scores are the share of available rubric points, not the share of tasks completed end to end. OpenAI reports 94.3% for Astra and 91.8% for Luna, plus 8.5 cents per task for Luna, 98% below Astra. OpenAI does not break down whether that figure includes retries, system integration, or human review.

ItemNumberBasis
Clinical safety99%325 tasks, reviewer judgment per task
Clinical helpfulness93%646 physician assessments, top two of five points
Admin score94.3% / 91.8%HealthAdminBench rubric points, Astra / Luna
Admin cost8.5 cents/taskGPT-6 Luna, maximum reasoning effort
Healthcare plugins21Official list, available in ChatGPT for Healthcare

The plugin list is Aiva, Amion, Assort Health, Availity Extend, Elation Health, Ferry, IMO Health MCP Server, Infinitus Relay, K Health, Luma Health, Metriport, Notable Integration Specs, Oracle HDI Overview, PenguinAI Medical Coding, Photon, Redox, Stedi, Subsalt, Trizetto, Turquoise, and UpToDate Expert AI. OpenAI describes them as integrations used inside ChatGPT for Healthcare and ChartGPT. They are a different capability from an agent that operates a screen; the plugin list is not evidence that a model can autonomously run an EHR.

On deployments, CVS Health uses agents to send documents, process referrals, and support specialist consults across clinical systems. One production workflow cleared a months-long document-sending backlog over a weekend; another handles the system steps for tens of thousands of consults a day, with human oversight, screenshots, and action records. Cigna Healthcare says case managers previously navigated more than 10 clinical and administrative systems and now pull the context together in seconds; the project went from concept to deployment in under eight weeks (HIT Consultant, Nov 16). Abridge is developing inpatient coordination, and Ambience’s agent stops when the record is ambiguous, asks the clinician, waits for approval before submitting a prior authorization, and keeps a step-by-step audit log.

On buying and compliance, organizations can purchase through sales, and ChatGPT for Healthcare is also available through an AWS Marketplace private offer; existing customers ask an administrator to enable computer use in ChatGPT Work and Codex. Workflows involving protected health information require services covered by a Business Associate Agreement plus the required security and retention settings. OpenAI says organization data is not used to train foundation models.

What others say

HIT Consultant frames the numbers as a shift in health-IT competition: at 8.5 cents per task, waiting for mature FHIR APIs is no longer the only option, and manual outsourcing and point-solution startups face margin pressure. It also argues that what OpenAI is really selling is governance, namely human review, step-by-step audit trails, and explicit decisions about which actions need approval.

iTMunch is more direct about the limits. These are evaluation results, not clinical outcomes; the demo uses synthetic data; HealthAdminBench scores rubric points, which is looser than full task success; and a signed BAA does not make every agent use automatically HIPAA-compliant. It also notes that models can still act without proper sign-off, a risk that deployment-side permissions and review have to absorb.

One wording difference is worth flagging. HIT Consultant calls the reviewers “independent reviewers,” while OpenAI’s own text says only “reviewers” and does not say whether they were external. We follow OpenAI’s wording: this is a vendor-run evaluation, not third-party certification.

Our take

What actually changes here is the cost structure of the last mile. Hospitals used to choose between waiting for vendors to ship an API and paying people to move data by hand. Computer-use agents offer a third option: pay a model to operate the interface that already exists. For hospital IT and revenue-cycle teams, three costs matter beyond the per-task model price: human review, maintenance when interfaces change, and retries plus audit overhead. The 8.5 cents covers the model run only, and OpenAI does not break down the rest.

For developers, computer use complements APIs rather than replacing them. Where an interface exists, use it; screen operation pays off only where no API exists or the API stops short. Reading the plugin list as evidence of autonomous EHR operation is the easiest mistake to make with this launch.

There are three honest problems. The 99% figure is a task-level safety rating over 325 vendor-designed tasks, with no real patient outcomes and no independent replication. Screen automation is brittle when software changes, and hospital software changes constantly, so maintenance will rise with interface churn. And liability for a wrong click still has to be answered by each deploying organization’s approvals, logs, and insurance arrangements.

What to watch

Three things are worth tracking: whether CVS and Cigna publish production metrics such as error rates, rework, and audit cost; whether an independent team reproduces the 99% and 8.5-cent figures; and whether EHR and payer-portal vendors restrict or standardize this kind of interface automation. Plugin adoption is the fourth signal: of the 21 names, which ones actually enter daily workflows, and which remain catalog entries.