OpenAI announced computer-use agents, improved clinical search and more than 20 new plugins for its healthcare products on November 16, 2026. The agents can navigate electronic health records, payer portals and other applications by reading screens, clicking and typing, extending ChatGPT into multi-step clinical and administrative work.
The release spans two companion announcements: the ChatGPT for Healthcare product update and OpenAI’s computer-use research and deployment report. Together, they describe available capabilities, proposed workflows, evaluation results and customer deployments.
For healthcare organizations, the practical distinction is between access and authorization. An available agent still needs approved applications, appropriate data protections and clear boundaries for human review. OpenAI’s published performance figures offer evidence for choosing a pilot, but they are vendor-reported results, not independent clinical validation.
What Is Available, and Through Which Product?
OpenAI says healthcare teams, developers and individual clinicians can access computer-use capabilities through ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API. Its plugin announcement is narrower: the new healthcare plugins are available in ChatGPT for Healthcare, including through ChartGPT, which brings ChatGPT into the EHR workflow.
| Capability | Announced access | Practical distinction |
|---|---|---|
| Computer-use agents | ChatGPT for Healthcare, ChatGPT for Clinicians and OpenAI API | Carry out multi-step work through application interfaces |
| Improved trusted clinical search | ChatGPT for Healthcare and ChatGPT for Clinicians | Find, evaluate and cite medical evidence |
| More than 20 new healthcare plugins | ChatGPT for Healthcare, including through ChartGPT | Connect healthcare tools within a conversation |
These are the availability claims in OpenAI’s November 16 posts. Product eligibility and workspace settings govern access; organizations should verify their enabled capabilities before designing a workflow around them.
The three components serve different purposes. Clinical search supplies evidence. Plugins connect named services. Computer use lets an agent operate software through its interface, including steps that existing APIs do not fully support.
I’d treat that last capability as a way to fill integration gaps, not a reason to abandon structured integrations. OpenAI’s CVS example specifically places computer use where APIs leave work unfinished. That makes a mixed approach, using integrations where available and screen interaction where necessary, a sensible deployment strategy.
The Plugin List Contains 21 Named Additions
OpenAI lists 21 new healthcare plugins. Its product announcement illustrates how several can work together:
- Visit preparation: Luma Health identifies appointments needing confirmation, while TriZetto checks insurance eligibility.
- Post-discharge follow-up: Metriport retrieves a discharge summary, while UpToDate supplies guidance relevant to medication changes.
- Inpatient coordination: Aiva retrieves a hospital’s insulin-pump policy, while Amion identifies the covering endocrinologist.
- MRI cost comparison: IMO Health identifies a procedure code, Stedi checks eligibility and benefits, and Turquoise supplies negotiated-rate information.
The other named additions are Assort Health, Availity Extend, Elation Health, Ferry Health, Infinitus Relay, KMED, Notable, Oracle HDI, PenguinAI, Photon, Redox and Subsalt. The clinical evidence plugin is listed as UpToDate Expert AI.
These examples are proposed workflows, not published tests of every integration. The announcement also describes preparing prescriptions for clinician review, patient outreach and operational analysis.
The MRI pricing example contains an important qualification: it asks ChatGPT to identify additional information needed to estimate the patient’s responsibility. A negotiated rate alone cannot answer that question. My view is that procurement demonstrations should preserve this distinction, especially when a generated comparison could influence where a patient receives care.
How the MRI Agent Workflow Works
OpenAI’s computer-use demonstration uses synthetic patient data. It follows a lower-back-pain message through two connected workflows:
- The agent reviews visit notes, physical therapy records and prior imaging in the EHR.
- It drafts an MRI order, an encounter note and a patient reply.
- The clinician reviews those drafts.
- The agent uses the same records to prepare a prior authorization request in the insurer’s portal.
- An authorization specialist reviews that request.
The consequential change is that the agent moves information between applications and prepares work inside them. It can reduce the manual transfer between a clinical record and a separate payer interface.
The demonstration also shows two review roles. The clinician reviews the clinical work; the authorization specialist reviews the payer submission. That division matters because checking clinical appropriateness and checking an authorization packet are different responsibilities.
For a pilot, I’d define review gates by action: preparing a draft, placing an order, sending a patient message and submitting a payer request should have explicit permissions. A broad instruction to “keep a human involved” leaves too much ambiguity about when the agent must stop. This is a deployment recommendation based on the demonstrated workflow, not a claim that OpenAI mandates the same approval sequence everywhere.
What the Safety and Cost Numbers Actually Measure
OpenAI evaluated GPT-6 Astra on EHR-inbox tasks shaped by physician input. Reviewers considered the agent’s actions alongside its notes, orders and patient messages, including whether inaccuracies, omissions or actions could plausibly cause patient harm.
Its published results use different denominators and scoring methods:
| Evaluation | OpenAI-reported result | What the measure represents |
|---|---|---|
| Astra clinical safety | 99% of 325 reviewed tasks rated safe | Task-level safety judgments |
| Astra helpfulness | 93% of 646 assessments rated very or extremely helpful | Ratings on a five-point scale |
| Astra on HealthAdminBench | 94.3% of available rubric points | Administrative-workflow scoring |
| Luna on HealthAdminBench | 91.8% of available rubric points | The same benchmark’s scoring |
| Luna benchmark cost | 8.5 cents per task, 98% below Astra | Reported cost with both models at maximum reasoning effort |
The 325 safety tasks and 646 helpfulness assessments are separate counts. They should not be merged into a single sample size.
OpenAI describes HealthAdminBench as an independent benchmark from Stanford researchers. That describes the benchmark’s origin; the Astra and Luna scores above come from OpenAI’s evaluation. A rubric score also differs from the percentage of workflows completed without intervention.
Luna trails Astra by 2.5 percentage points of available rubric points while carrying a much lower reported cost. That is a useful reason to evaluate Luna for bounded administrative tasks. It leaves a separate question: whether the missing points affect harmless formatting details or consequential omissions in a particular organization’s workflow.
Authorization Errors Deserve Separate Attention
Fierce Healthcare’s report adds context to the tasks outside the safe category. OpenAI’s Karan Singhal told the publication that researchers had not identified cases creating a specific, plausible risk of patient harm, while acknowledging:
Frequently Asked Questions
3 questions
1Are the New Plugins Available in ChatGPT for Clinicians?
OpenAI announces the plugins for ChatGPT for Healthcare, including through ChartGPT. It separately announces computer-use agents and improved clinical search for ChatGPT for Clinicians.