LATEST
AI Tools

ChatGPT for Healthcare Gets Computer-Use Agents, 21 Plugins and a 99% Safety Claim

Updated November 17, 2026
17 min read
Editor's Note: This article was human-reviewed for accuracy
A female doctor in a white coat with a stethoscope uses a laptop while writing notes at a wooden table. (Photographer: Pavel Danilyuk, Pexels photo 5998475.)

Quick Answer & Key Takeaways

  • • OpenAI added computer-use agents to ChatGPT for Healthcare on November 16, 2026, letting AI click, type and move between electronic health records and payer portals to handle referrals, prior authorizations and discharge planning.
  • • The same update brings stronger clinical search and 21 named healthcare plugins, and OpenAI reports GPT-6 Astra was rated clinically safe on 99% of 325 reviewed EHR tasks in its own evaluation.
  • • Organizational pricing is not published, but ChatGPT for Clinicians remains free for eligible, verified clinicians in the United States.

Ask a hospital team what slows care down and the answer is rarely medicine. It is the retyping: a clinician recommends an MRI, the details an insurer needs are already in the patient's record, and someone still has to pull them out and enter them into a separate portal. On November 16, 2026, OpenAI announced that ChatGPT for Healthcare will now do that moving for itself. The update adds computer-use agents that click, type and move between the applications care teams already use, alongside stronger clinical search and 21 named healthcare plugins. It arrives with safety numbers that need reading carefully: GPT-6 Astra was rated clinically safe on 99% of 325 reviewed EHR inbox tasks in the company's own evaluation. This article explains what the agents can do, what the numbers do and do not prove, and what it costs.

What launched on November 16

OpenAI published two posts that work as one launch. The first covers computer-use agents for healthcare teams; the second covers the wider ChatGPT for Healthcare update. Together they contain three parts. Computer-use agents navigate existing healthcare software by reading screens, clicking, typing and moving between applications, which lets them carry out multi-step work where no API connects the systems involved. Trusted clinical search, the way ChatGPT finds, evaluates and cites medical evidence, has been upgraded in ChatGPT for Healthcare and in ChatGPT for Clinicians. And the healthcare plugin ecosystem grows by more than 20 new plugins, available in ChatGPT for Healthcare and inside EHR workflows through ChartGPT, the version of ChatGPT that sits within the EHR layout.

The agents are available now through the OpenAI API, ChatGPT for Healthcare and ChatGPT for Clinicians. Healthcare organizations that already use the product can ask a workspace administrator to switch computer use on in ChatGPT Work and Codex. Individual clinicians in the United States can sign up for a free ChatGPT for Clinicians account, subject to verification as a clinician. Developers can build with the capability through the OpenAI API and OpenAI's computer-use guide.

The timing matters because the product is not new. ChatGPT for Healthcare launched on January 8, 2026, as the enterprise tier of OpenAI's healthcare suite, with early customers including Boston Children's Hospital, Cedars-Sinai, HCA Healthcare and UCSF, according to launch reporting. OpenAI added an Epic integration on September 1, 2026, which brought authorized patient context into ChatGPT. The November 16 update is the step from answering questions about care to doing the administrative work around it. OpenAI says more than 230 million people already ask ChatGPT health questions every week, which is the audience this governed, enterprise version is built to serve without mixing consumer chats into it.

How the agents work

A computer-use agent does what a person would do at a keyboard, only under instruction. OpenAI defines it plainly: the agents carry out multi-step workflows by clicking, typing and moving between applications. In healthcare that means opening the EHR, reading the chart, drafting the documents, then moving to an insurer's portal to prepare the request, instead of a staff member copying facts from one window to another.

OpenAI's demonstration, built on synthetic patient data, shows the shape of it. A patient messages a clinician about lower back pain. The agent reviews the patient's 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 same records are then used to prepare a prior authorization request in the insurer's portal, for an authorization specialist to review. Nothing is submitted on its own. The drafts wait for a person, which is the design point OpenAI keeps returning to: the clinician and the authorization specialist stay in control of the workflow, and the agent moves information between its steps.

The same pattern applies to the administrative tasks the announcement names: referrals, prior authorizations, discharge planning and outreach between visits. Agents doing real work inside existing systems is a wider pattern now: Walmart already lets AI agents shop on customers' behalf, with spending limits built in. Karan Singhal, OpenAI's head of Health AI, told Fierce Healthcare that the company set a research bet about six months ago that computer use would become reliable enough for healthcare workflows, and that this release is the result. The pitch is not diagnosis. It is paperwork, and the test is whether an agent can be trusted to move it without inventing or skipping a step.

The safety numbers, and what they do not prove

OpenAI published two sets of measurements with the launch, and both need reading as the company's own evaluations rather than independent results.

On clinical work, OpenAI evaluated GPT-6 Astra on tasks clinicians handle in their EHR inboxes. The tasks were shaped with physician input and require multi-step work: clinical reasoning over a patient's history plus navigation of complex interfaces. Reviewers examined the model's actions alongside the notes, orders and patient messages it produced, judging whether the work would be helpful and clinically safe. Clinical safety meant checking whether inaccuracies, omissions or actions could plausibly lead to patient harm. Overall, 99% of 325 reviewed tasks were rated safe, and physicians rated the work "very helpful" or "extremely helpful" in 93% of 646 assessments.

The 1% that was not rated safe deserves the same attention as the 99%. Singhal told Fierce Healthcare that researchers found no case in that group where the model created a specific, plausible risk of patient harm, but that in some cases the model might overreach and take an action without getting the proper sign-off or authorization. That is the failure mode to watch: not a wrong diagnosis, but an agent acting before a person has approved it. His own summary was blunt: "The models are not perfect today, and what we're sharing is an update on research results where we're seeing this critical point where the models seem to be useful in practice."

On administrative work, OpenAI used HealthAdminBench, a benchmark from Stanford researchers that measures performance on common administrative workflows, including prior authorizations, claim appeals and durable medical equipment orders. The benchmark contains 135 tasks across three difficulty levels. GPT-6 Astra earned 94.3% of available rubric points and GPT-6 Luna earned 91.8%, with both models run at maximum reasoning effort. The cost gap is the striking part: Luna cost 8.5 cents per task, 98% less than Astra on the same benchmark. That works out to roughly $4.25 per task for Astra, by simple arithmetic on OpenAI's figures, against about $0.085 (roughly 6p, or about C$0.12 at November 17, 2026 market rates) for Luna. For high-volume administrative work, the cheaper model doing nearly as well is the number that decides whether the economics work.

MeasurementResultWhat it is
EHR inbox tasks rated clinically safe (GPT-6 Astra)99% of 325 reviewed tasksOpenAI's own evaluation, physician reviewers
Work rated "very helpful" or "extremely helpful"93% of 646 assessmentsOpenAI's own evaluation, five-point scale
HealthAdminBench, GPT-6 Astra94.3% of rubric pointsStanford researchers' benchmark, 135 tasks
HealthAdminBench, GPT-6 Luna91.8% of rubric pointsSame benchmark, maximum reasoning effort
Cost per administrative task, GPT-6 Luna$0.085 (about 6p, about C$0.12)OpenAI-reported benchmark cost, not a list price
Cost gap, Luna vs Astra98% lessOpenAI-reported; implies roughly $4.25 per Astra task

Two caveats belong with every number in that table. First, the safety and helpfulness figures are physician assessments of the model's work, not measured patient outcomes. No one has shown that these agents change what happens to patients, only that reviewers judged their draft work safe and helpful. Second, the per-task cost is a benchmark cost at maximum reasoning effort, not a hospital's price per completed case. A real deployment adds implementation, partner subscriptions, infrastructure, security review, human checks and exception handling, none of which OpenAI prices. Treat 8.5 cents as a unit-economics signal, not an invoice.

What it costs

There is no published price for ChatGPT for Healthcare as an organization. OpenAI says organizations can buy it by contacting its sales team or through a private offer on AWS Marketplace, and the company has published no seat price, tier list or per-task rate for the product itself. That absence is the honest answer to the pricing question, and it should be stated plainly: any article quoting a ChatGPT for Healthcare price is guessing.

What is published is the split between the two products. ChatGPT for Clinicians is free for eligible, verified clinicians in the United States, for reviewing medical evidence and preparing clinical documentation. It includes improved clinical search and a selection of the new plugins, while the full plugin set is available in the organizational product. Developers using the capability through the OpenAI API pay the API's own rates for the models involved, which are separate from the healthcare products. GPT-6 Astra and GPT-6 Luna are the models behind the agents; their API pricing is a different question from the cost of the healthcare workspace, and OpenAI has not published a per-task price for the agents themselves.

That split shapes who this update is for in practice. A verified US clinician can try the agents and the improved search now, free, in ChatGPT for Clinicians. A hospital or payer faces a sales conversation, a workspace configuration, and the cost of connecting its own systems, which is where the real bill lives. The 8.5 cents per task for Luna is useful as a comparison between models, not as a budget line.

The 21 plugins, and how they chain together

The plugin list is the most concrete part of the update, because each plugin names a real tool and a real job. OpenAI lists 21: 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. All are available in ChatGPT for Healthcare, including through ChartGPT. Workspace administrators must switch them on; individual clinicians get a selection of them in ChatGPT for Clinicians.

The value is in chaining them inside one conversation, and OpenAI's examples show how that works. A primary care physician can pull a discharge summary from another hospital through Metriport, have ChatGPT identify medication changes, and draw on UpToDate Expert AI for recommended follow-up monitoring, with links to both sources. An inpatient nurse admitting a patient who uses an insulin pump can combine Aiva's hospital-policy information with Amion's on-call directory to find the covering endocrinologist. For cost and access, IMO Health identifies the procedure code for a knee MRI without contrast, Stedi checks insurance eligibility and benefits, and Turquoise Health supplies the plan's negotiated rate for that code at two hospitals, with the workflow also flagging what else is needed to estimate what the patient would owe. Luma Health and Trizetto build pre-visit confirmation lists; Assort Health and Ferry Health combine referral trends with outstanding scheduling work.

GroupPluginsWhat they do
Clinical reference and terminologyUpToDate Expert AI, IMO Health MCP ServerEvidence, follow-up monitoring, procedure codes
EHR and interoperabilityElation Health, Redox, Metriport, Oracle HDI OverviewRecords, discharge summaries, data exchange
Revenue cycle and clearinghouseAvaility Extend, Stedi, Trizetto, PenguinAI Medical Coding, TurquoiseEligibility, benefits, coding, negotiated rates
Engagement and pharmacyAssort Health, Luma Health, Photon, Infinitus Relay, Amion, Aiva, Ferry, K Health, Notable Integration Specs, SubsaltScheduling, referrals, outreach, benefits verification

Two limits sit under that list. First, every plugin is a separate connection, and access depends on the workspace, administrator settings, connected services and, in some cases, separate partner accounts. A plugin being listed does not mean it is switched on for a given team. Second, the plugins join a broader ecosystem, and OpenAI is inviting more healthcare companies to build for it, so the list will keep changing. The right question for a team evaluating this is not how many plugins exist but which of its own workflows the enabled set actually covers.

Who is already using it

OpenAI named four organizations building on the capability, and the examples are more specific than the usual launch quotes. CVS Health is the furthest along. One production agent cleared a months-long document-sending backlog over a weekend, and another handles the system steps behind tens of thousands of specialist consults a day, including steps that APIs do not fully support. CVS uses validation checks and human oversight, with screenshots and action records to verify what the agent did, and the data stays inside CVS Health's enterprise-controlled infrastructure under its AI governance team. CVS is also open-sourcing Agent Kit, built with the OpenAI Agents SDK, as a repeatable process other organizations can use.

Cigna Healthcare is using the models to assemble member context that case managers used to gather from more than 10 clinical and administrative systems, including benefits, coverage, providers, appointments and prior interactions. OpenAI says the team moved from concept to deployment in under eight weeks, and that computer use can now bring that information together in seconds. Abridge is developing agents for daily applications that lack APIs, including a planned inpatient use case where a clinician can ask for help with discharge barriers while agents work at the same time on medication access and home-health coordination, with the clinician brought in when a decision needs clinical judgment. Ambience built its example around control: its agent prepares a prior-authorization request from the chart, stops when the record is ambiguous, and waits for approval before submitting, while health systems set which actions the agent can take, can pause it at any point, and get a step-by-step audit log.

Smaller companies are testing the same pattern. Assort Health used computer use in Codex to close a referral loop that Athena's API does not cover, selecting the referring provider, entering the referral details and marking the referral "visit scheduled" so it dropped off the staff follow-up list. Fierce Healthcare also names Bunkerhill, Forus, Gravity and Tennr as companies testing agents that interact directly with EHRs and payer portals for referral management, authorization checks, medication approvals and other operational tasks. These are OpenAI's accounts of its partners' work, not independent verification, and the production cases are the ones to watch: a weekend backlog cleared at CVS is a result; a demo is a promise. The same agent push is visible elsewhere in OpenAI's work: its Daybreak agents do penetration testing and patching, another case of agents moving from chat into systems that used to need a person at the keyboard.

Clinical search gets better, on paper

The second half of the launch is quieter but affects every user. Trusted clinical search in ChatGPT for Healthcare and ChatGPT for Clinicians now finds, evaluates and cites medical evidence differently, with citations clinicians can inspect and practical next steps attached to the evidence. OpenAI measured the change with a blind evaluation in which physicians reviewed responses to real-world clinical questions across specialties, including multi-turn conversations. Across more than 1,300 reviews, physicians rated the updated experience more highly on all eight dimensions OpenAI measured: factual correctness, safety, actionability, calibration, source authoritativeness, locality alignment, temporal alignment and goal satisfaction.

HIT Consultant, reporting OpenAI's evaluation, puts numbers on three of those dimensions: factual correctness rose from 93.4% to 97.6%, clinical safety from 89.4% to 95.5%, and temporal alignment from 94.4% to 98.8%. Those are physician assessments of answers, not patient outcomes, and the evaluation compared the previous and updated ChatGPT for Clinicians experiences rather than testing against a rival product. Still, the direction is consistent across all eight dimensions, which is a stronger signal than a single headline figure. For clinicians, the practical change is that answers should come with evidence they can check, including local and current sources, rather than a confident paragraph with no trail back to it.

How it compares

OpenAI is not the only company putting a general-purpose assistant into healthcare workflows. Google Cloud launched Gemini for healthcare in preview in the United States on November 10, 2026, six days before OpenAI's announcement. Anthropic launched Claude for Healthcare in January 2026, and Amazon launched Amazon Connect Health on March 5, 2026. The products overlap in audience but differ in emphasis: OpenAI's pitch is a cited assistant that clinicians and staff talk to, while Google's is designed to go into claims, eligibility and EHR systems and assemble finished work such as a prior authorization packet. That distinction is the one to keep in mind when comparing them. OpenAI's November 16 update narrows the gap on the assembly side, because its agents now do the clicking through EHRs and payer portals that Google's product was built around, but the two are still sold and governed differently, and neither has published organizational pricing that would let a buyer compare them on cost. For readers weighing options, the honest comparison is on which workflows each product actually covers in a given hospital's systems, not on launch dates.

Privacy, control and the HIPAA question

Healthcare is the field where an agent's mistakes carry the most consequence, so the controls matter as much as the capability. OpenAI's position, stated in both posts, is that organization data in ChatGPT for Healthcare is not used to train its foundation models, and that content shared with the product is excluded from training. The product provides role-based access controls and audit logs, and organizations can build additional review and approval steps into their workflows and configure which applications and environments agents can use, limiting access to the records and actions each task needs.

The HIPAA position needs stating carefully. Both healthcare products offer a Business Associate Agreement with OpenAI to support HIPAA-compliant workflows, but a BAA is a starting condition, not a guarantee. Organizations must use eligible services covered by their BAA, apply the required security and retention settings, and extend those protections to connected systems. The consumer product, ChatGPT Health, is a different product and does not carry the same compliance posture. Anyone evaluating the enterprise product should confirm which services their BAA covers before putting protected health information through an agent, and should treat the audit log as a requirement rather than a feature: if a workflow cannot show what the agent did and who approved it, it is not ready for production use.

Common mistakes to avoid

The first mistake is reading the safety numbers as a verdict on patient outcomes. They are physician assessments of draft work in an evaluation OpenAI designed and ran. The 1% of tasks not rated safe, where the model might act without proper sign-off, is the number that decides whether a deployment is governed well enough, and it will only be tested by real use under audit.

The second is assuming the plugins are on. Workspace administrators must enable them, individual clinicians get only a selection, and some workflows need separate partner accounts. A team that buys the product and never configures the plugin set has bought a chat window, not a workflow tool.

The third is budgeting from the benchmark. The 8.5 cents per Luna task is a benchmark cost at maximum reasoning effort, not a price, and the organizational product has no published price at all. The real cost is the sales agreement, the workspace setup, the partner subscriptions and the human review time, and none of those are in OpenAI's announcement.

The fourth is confusing the two healthcare products. ChatGPT for Healthcare is the governed organizational workspace; ChatGPT for Clinicians is the free account for verified US clinicians; ChatGPT Health is the consumer experience. They share a brand and little else in compliance terms.

Frequently asked questions

What is ChatGPT for Healthcare?

It is OpenAI's enterprise healthcare product: a governed ChatGPT workspace for clinical, research and administrative teams, launched January 8, 2026. It offers a Business Associate Agreement, role-based access controls and audit logs, and it now includes computer-use agents, upgraded clinical search and more than 20 healthcare plugins.

How much does ChatGPT for Healthcare cost?

OpenAI has not published a price. Organizations buy it by contacting OpenAI's sales team or through a private offer on AWS Marketplace. ChatGPT for Clinicians, the separate product for individual clinicians, is free for eligible, verified clinicians in the United States.

What can the computer-use agents actually do?

They navigate existing software by reading screens, clicking, typing and moving between applications, to handle referrals, prior authorizations, discharge planning and outreach between visits. They draft orders, notes and requests for a clinician or specialist to review, rather than submitting work on their own.

Is patient data used to train OpenAI's models?

According to OpenAI, no. Organization data in ChatGPT for Healthcare is not used to train its foundation models, and content shared with the product is excluded from training. Production data at CVS Health, for example, stays inside CVS's enterprise-controlled infrastructure.

Is ChatGPT for Healthcare HIPAA compliant?

It can support HIPAA-compliant workflows, but compliance is not automatic. Both healthcare products offer a Business Associate Agreement with OpenAI, and organizations must use eligible services covered by that agreement, apply the required security and retention settings, and extend protections to connected systems.

How safe are the agents, according to OpenAI?

In OpenAI's own evaluation, reviewers rated 99% of 325 GPT-6 Astra EHR inbox tasks clinically safe, and physicians rated the work very or extremely helpful in 93% of 646 assessments. On the Stanford HealthAdminBench benchmark, Astra earned 94.3% and Luna 91.8% of rubric points. These are company-run evaluations, not independent results or measured patient outcomes.

OpenAI's healthcare update is a genuine step: agents that do the clicking through systems no API connects, a plugin list that covers real workflows, and safety numbers published with their limits visible. The open questions are the ones no launch post can answer: what the organizational product costs, how the agents behave under audit in a hospital that is not a named partner, and whether the 1% of tasks that were not rated safe stays a rounding error or becomes the story. For a verified US clinician, the free Clinicians account is the low-risk way to find out. For a health system, the next step is a controlled pilot with the audit log switched on and a human in every approval step. This is general information, not medical advice.

Sources

17 min read
93 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more