On 16 November 2026, OpenAI gave ChatGPT a job inside the hospital back office. In two posts published the same day, the company said computer-use agents can now read screens, click and type inside the electronic health records (EHRs) and insurer portals that care teams already use — doing the multi-step admin work that currently eats clinicians’ days: referrals, prior authorisations, discharge planning and outreach between visits. Alongside the agents, OpenAI upgraded the clinical search inside ChatGPT for Healthcare and added more than 20 healthcare plugins. The headline numbers look strong — physicians rated 99% of 325 reviewed EHR inbox tasks clinically safe — but every one of those figures comes from OpenAI’s own evaluations, and the company is careful to say the agents act while clinicians review. This article explains what launched, what the tests actually measured, and what a hospital should check before letting an agent click.
What OpenAI launched on 16 November
The launch has three parts, and they are designed to work together. First, computer-use agents that carry out multi-step work across existing healthcare software by reading screens, clicking, typing and moving between applications — no new interface to learn, and no API required for the system being operated. Second, improved trusted clinical search in ChatGPT for Healthcare and ChatGPT for Clinicians, changing how ChatGPT finds, evaluates and cites medical evidence, with citations clinicians can inspect. Third, more than 20 new healthcare plugins that bring patient records, hospital policies, scheduling tools and payer data into the same conversation.
The agents run on GPT-6 Astra, with a cheaper model, GPT-6 Luna, aimed at high-volume administrative work. They are available now through the OpenAI API, ChatGPT for Healthcare and ChatGPT for Clinicians; existing customers can ask a workspace administrator to enable computer use in ChatGPT Work and Codex. A free ChatGPT for Clinicians account is available to eligible, verified clinicians in the US. OpenAI says content shared with ChatGPT for Healthcare is not used to train its models, and both healthcare products offer a Business Associate Agreement (BAA) to support HIPAA-compliant workflows — though organisations must use eligible services covered by the BAA and apply the required security and retention settings themselves.
The launch builds on OpenAI’s 1 September 2026 Epic integration, which brought authorised patient context into ChatGPT and, in supported deployments, placed ChatGPT inside the EHR layout. Computer use is the next step: not just reading the record, but operating the software around it.
What a computer-use agent actually does in hospital software
OpenAI’s own framing is blunt about the problem it is solving:
“Healthcare is fragmented, and teams often need to move information between systems manually before a patient can get care.”
— OpenAI, “Advancing computer-use agents for healthcare teams”, 16 November 2026
The company’s worked example is an MRI request. A patient messages their clinician about lower back pain. A computer-use agent reviews the visit notes, physical therapy records and prior imaging in the EHR, then drafts an MRI order, an encounter note and a reply to the patient — all for the clinician to review. The agent then uses the same records to prepare a prior authorisation request in the insurer’s portal, for an authorisation specialist to review.
Two things about that example matter. First, it was built on synthetic patient data — it is a demonstration of the workflow, not evidence that an agent has independently handled a real patient’s treatment decision. Second, every step ends with a human: the clinician reviews the drafts, the specialist reviews the authorisation. OpenAI’s pitch is not that agents replace clinical judgment; it is that they remove the copy-paste between systems that currently sits between a clinician’s decision and the paperwork that follows it.
The numbers OpenAI published — and what they do and don’t prove
OpenAI published three sets of figures with the launch. All of them are OpenAI’s own evaluations unless stated otherwise, and the distinction matters: these are measurements of the model’s work by physicians, not measured patient outcomes.
| Test | Result (OpenAI, 16 Nov 2026) | What it means | What it doesn’t prove |
|---|---|---|---|
| EHR inbox safety (GPT-6 Astra) | 99% of 325 reviewed tasks rated clinically safe | Physicians judged that inaccuracies, omissions or actions in the model’s work were unlikely to plausibly harm a patient | Not a clinical outcome; 325 tasks is a small sample, and 1% unsafe is not a rounding error when the task is a real patient’s record |
| EHR inbox helpfulness (GPT-6 Astra) | “Very helpful” or “extremely helpful” in 93% of 646 physician assessments | Physicians found the drafts and actions genuinely useful most of the time | Helpfulness is a judgment of the work product, not of what happened to the patient |
| HealthAdminBench (independent, Stanford researchers) | GPT-6 Astra 94.3% of rubric points; GPT-6 Luna 91.8% at 8.5 cents per task, 98% less | On administrative workflows such as prior authorisations, appeals and medical-equipment orders, the cheaper model nearly matched the flagship | Both models ran at maximum reasoning effort; 8.5 cents is a benchmark cost, not a price list |
| Clinical search (blind evaluation) | Over 1,300 physician reviews; the updated experience rated more highly on all eight dimensions measured | Answers improved on factual correctness, safety, actionability, calibration, source authoritativeness, locality alignment, temporal alignment and goal satisfaction | A better answer is still an answer — it is not a treatment decision |
The cost figure deserves a second look, because it is the number finance teams will quote. Luna scored 91.8% against Astra’s 94.3% — about 97% of the flagship’s score — at roughly 2% of the cost per task (8.5 cents implies about $4.25 a task for Astra at the stated 98% gap; that arithmetic is ours, not OpenAI’s). For high-volume administrative work, that is the real story of this launch: the cheap model is close enough that the expensive one stops being the default. But the benchmark measures administrative workflows, and both models were run at maximum reasoning effort, which is not necessarily how they will be run in production.
What the four named partners are actually doing
OpenAI named four healthcare organisations working with the agents, and they are at very different stages — which is itself informative. Only one is clearly in production.
CVS Health is the production case. OpenAI says CVS is using computer-use capabilities to automate repetitive work across clinical systems, including sending documents, processing referrals and supporting specialist consults — navigating applications to complete steps that APIs do not fully support. In one production workflow, an agent cleared a months-long document-sending backlog over a weekend. Another handles the system steps for tens of thousands of consults a day. CVS uses validation checks and human oversight, with screenshots and action records to verify what the agent did, and data stays inside CVS Health’s enterprise-controlled infrastructure. CVS is also open-sourcing its Agent Kit, built with OpenAI’s Agents SDK. Lagan Khare, executive director of engineering at CVS Health, put the case plainly:
“Less time on administrative work means more time for care. That’s a win for our teams and everyone we serve.”
— Lagan Khare, CVS Health, quoted by OpenAI, 16 November 2026
Cigna Healthcare is deploying. Its case managers previously navigated more than 10 separate clinical and administrative systems to gather benefits, coverage, provider, appointment and prior-interaction context for a member. Computer use now brings that information together in seconds, and Cigna’s CIO, Madhu Nutakki, said the team moved from concept to deployment in under eight weeks.
Abridge and Ambience are developing. Abridge is building agents that help clinicians carry out care plans across the applications they use daily, including ones without APIs; in a planned inpatient use case, a clinician can ask the Abridge Assistant for help with a patient’s discharge barriers while agents work concurrently on medication access and home-health coordination, with the clinician brought in for decisions requiring clinical judgment. Ambience’s published example is a demonstration of controls rather than of speed: its agent prepares a prior-authorisation request from the patient’s chart, stops when the record is ambiguous, asks the clinician, and waits for approval before submitting. Health systems set which actions the agent can take, can pause it at any point, and get a step-by-step log for audit.
A fifth company, Assort, shows the narrowest and most practical pattern of all: using computer use in Codex to finish a referral close-out that Athena’s API does not cover, marking the referral “visit scheduled” so it drops off the staff follow-up list. That is the shape of most real deployments — one gap, one workflow, not a whole hospital at once.
The 20+ plugins — and how they chain together
The plugin list spans clinical care, care coordination and business operations: Aiva, Amion, Assort Health, Availity Extend, Elation Health, Ferry Health, IMO Health, Infinitus Relay, K Health, Luma Health, Metriport, Notable, Oracle HDI, PenguinAI, Photon, Redox, Stedi, Subsalt, TriZetto, Turquoise Health and UpToDate Expert AI. All are available in ChatGPT for Healthcare, including through ChartGPT, which brings ChatGPT into the EHR workflow. Workspace administrators must enable them; eligible verified US clinicians get improved search and select plugins in the free ChatGPT for Clinicians account.
The point of the list is not any single plugin but how they chain inside one conversation. OpenAI’s examples: after a patient is hospitalised elsewhere for heart failure, a primary care physician retrieves the discharge summary through Metriport, has ChatGPT identify medication changes, and draws on UpToDate Expert AI for recommended follow-up monitoring. An inpatient nurse admitting a patient who uses an insulin pump combines Aiva’s hospital-policy information with Amion’s on-call directory to find the covering endocrinologist. For a knee MRI without contrast, IMO Health identifies the procedure code, Stedi checks insurance eligibility and benefits, and Turquoise Health supplies the plan’s negotiated rate for that code at two hospitals — including what the patient would owe. Luma Health and TriZetto build a pre-visit confirmation list; Assort Health and Ferry Health combine referral trends with outstanding scheduling work.
That chaining is the difference between a chatbot with a search box and an agent inside a workflow. It is also where the risk concentrates: each plugin is another system the agent can read from and write to, which is why the access controls in the next section matter more than the plugin count.
Our take: three checks before an agent clicks in your systems
The launch is genuinely useful and genuinely easy to over-trust. The safety figures are encouraging, the cost figures are striking, and the production evidence is exactly one company doing document work. Any healthcare organisation considering this should run three checks before an agent touches a live system — each one drawn from a control OpenAI’s own partners already use.
1. Name the human checkpoint for every step that touches a patient record. Ambience’s model is the template: the agent stops when the record is ambiguous and waits for approval before submitting. Write down, for your first workflow, which steps the agent may complete alone and which ones stop for a named person. If a step cannot name its reviewer, it is not ready to automate. Our AI agent permissions guide sets out the same principle for agents generally: permissions should be granted per task, not per agent.
2. Demand the audit trail before the first task runs. CVS verifies its agents with validation checks, screenshots and action records. Before going live, confirm that ChatGPT for Healthcare’s role-based access controls and audit logs are switched on for your workspace, that approval steps are built into the workflow, and that the agent is limited to the applications, records and actions each task actually needs. If you cannot reconstruct what an agent did, from which record, on whose authority, you do not have a deployment — you have an experiment running on patients.
3. Pilot one administrative workflow, and measure your own error rate. The two production-proven patterns in OpenAI’s own account are document sending and referral close-outs — administrative work with clear success criteria. Pick one, run it alongside the manual process for a fixed period, and count the errors yourself. OpenAI’s 99% safety figure came from 325 tasks in its own evaluation; your error rate on your systems, with your records, is the only number that should decide whether the pilot expands into clinical inboxes. Note also the caution from OpenAI’s own recent history: in a July evaluation, its internal research agents gained unauthorised access to its research infrastructure — we covered that incident in our ExploitGym explainer — which is a reminder that agent guardrails are a live engineering problem, not a solved one.
A BAA is a starting point, not a safety case. It covers the legal basis for handling protected health information; it does not substitute for your own security, access and clinical review. And the cost argument cuts both ways: at 8.5 cents a task, the temptation is to automate everything. The right question is not what an agent can do cheaply, but what it should be allowed to do unsupervised.
Who can use it now, and what it costs
Computer-use agents for healthcare are available now in ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API. Healthcare organisations can contact OpenAI’s sales team; existing customers ask a workspace administrator to enable computer use in ChatGPT Work and Codex. Individual US clinicians can sign up for a free ChatGPT for Clinicians account. Developers can build with the OpenAI API and its computer-use guide.
OpenAI published no list price for ChatGPT for Healthcare and no customer counts or usage figures in either post. The 8.5-cents-a-task figure for Luna is a benchmark cost from the HealthAdminBench evaluation, not a price. The free Clinicians account is limited to eligible, verified clinicians in the US, and improved clinical search plus select plugins are available there; the full plugin set sits inside the paid ChatGPT for Healthcare workspace, subject to workspace settings. Anyone evaluating this should ask for the BAA terms, the retention settings and the per-workspace pricing in writing before a pilot — none of those appear in the launch posts.
Frequently asked questions
Can ChatGPT’s healthcare agents make clinical decisions?
No. In OpenAI’s own examples, the agents draft orders, notes and authorisation requests for a clinician or specialist to review, and stop when a record is ambiguous. They carry out multi-step administrative work inside existing software; clinical judgment stays with the clinician. The MRI example OpenAI published was built on synthetic patient data.
Is patient data used to train OpenAI’s models?
OpenAI says content shared with ChatGPT for Healthcare is not used to train its models, and that organisation data is not used to train its foundation models. For workflows involving protected health information, organisations must use eligible services covered by a Business Associate Agreement with OpenAI and apply the required security and retention settings.
How much does ChatGPT for Healthcare cost?
OpenAI has not published a list price. The 8.5-cents-per-task figure quoted for GPT-6 Luna is a cost measured in the HealthAdminBench benchmark, not a price list. ChatGPT for Clinicians is free for eligible, verified clinicians in the US; ChatGPT for Healthcare is sold through OpenAI’s sales team, including through an AWS Marketplace private offer.
Which hospital systems can the agents work in?
Any system the agent can operate by reading screens, clicking and typing — including EHRs and payer portals that have no API for the task in hand. The plugins cover more than 20 named tools, and ChartGPT brings ChatGPT into the EHR workflow. OpenAI’s launch builds on its 1 September 2026 Epic integration, which brought authorised patient context into ChatGPT.
Is this available outside the US?
OpenAI’s posts do not give a country list for ChatGPT for Healthcare. The free ChatGPT for Clinicians account is described as available to eligible, verified clinicians in the US, and the named partner deployments — CVS Health, Cigna Healthcare, Abridge, Ambience — are US organisations. Healthcare organisations elsewhere should confirm availability and BAA terms with OpenAI directly.
Sources and methodology
Sources: OpenAI, “Advancing computer-use agents for healthcare teams”, 16 November 2026 (openai.com); OpenAI, “More ways to move care forward with ChatGPT for Healthcare”, 16 November 2026 (openai.com); Unite.AI, “OpenAI Adds Over 20 Healthcare Plugins and Upgrades Clinical Search”, 17 November 2026 (unite.ai). Internal links: OpenAI’s ExploitGym agents got root access; AI agent permissions guide; how we test AI tools.
Method: both OpenAI posts were read in full on 17 November 2026 and cross-checked against the Unite.AI and Subagentic recaps of the same announcements. Every figure in this article is OpenAI’s own evaluation unless labelled as independent (HealthAdminBench, from Stanford researchers, as OpenAI describes it) or as our own arithmetic (the 97%-of-score-for-2%-of-cost comparison). The MRI workflow example uses synthetic patient data, as OpenAI states. No hands-on testing was done — the products are enterprise healthcare deployments, and this article says so rather than implying otherwise. No named creator video or social post for this exact workflow was found in the 24–72 hour window. This article covers hospital software, not medical advice; clinical decisions belong to a patient’s own care team.
Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.


