Latest
AI Affairs, home

Monday 16 November 2026

Technology

OpenAI adds computer-use agents and healthcare plugins to ChatGPT

The release also updates clinical search, as OpenAI reports safety and usefulness ratings for GPT-6 Astra on reviewed electronic health record tasks.

OpenAI’s headquarters building at 1515 Third Street, with palm trees and overhead wires
Photo: Coolcaesar, CC BY 4.0, via Wikimedia Commons (cropped)

OpenAI introduced computer-use agents, improved clinical search and more than 20 plugins for ChatGPT for Healthcare on 16 November. The agents are designed to carry out work across existing healthcare software by reading screens and operating applications, extending the product beyond finding information and drafting answers.

Key points

  • OpenAI says its agents can move between healthcare applications, with controls over access and steps requiring human review.
  • In OpenAI’s evaluation, 99% of 325 reviewed GPT-6 Astra inbox tasks were rated clinically safe.
  • More than 20 new healthcare plugins are available in ChatGPT for Healthcare, alongside an update to clinical search.

The announcement came at HLTH26, a healthcare technology conference in Las Vegas, Fierce Healthcare reported. OpenAI is working with CVS Health, Cigna Healthcare, Abridge and Ambience on applications that include referrals, prior authorisations, discharge planning and preparation for care managers.

Agents cross EHRs and payer portals

A computer-use agent works through a software interface much as a person would: it reads what is on the screen, clicks, types and moves to another application when the task requires it. In healthcare, that could connect steps spread across an electronic health record (EHR), scheduling software and an insurer’s portal. OpenAI says organisations can choose which applications an agent may access and which actions need human review, with audit logs recording what the agent did.

Requesting an MRI could involve carrying details from a medical record into a separate insurance request. Moving those details across screens could become less manual, with a proposed order and authorisation request left for review.

In a research demonstration using synthetic patient data, an agent examined visit notes, physical therapy records and earlier imaging, then drafted an MRI order, an encounter note and a patient reply for review. It also prepared a prior authorisation request in an insurer’s portal for a separate review. The demonstration illustrates the movement between applications, rather than an unattended clinical decision.

OpenAI says CVS Health is using computer-use capabilities for document sending, referrals and specialist consults. CVS Health said an agent cleared a months-long document-sending backlog in one production workflow over a weekend, Fierce Healthcare reported. Abridge is building agents intended to help carry out care plans across applications, including ones without application programming interfaces, or APIs. An agent that operates the existing screen can work through a gap between systems without relying on an integration that covers every step.

OpenAI had previously released GPT-6.1 Terra for computer use at lower API prices. The healthcare work brings that kind of interaction into clinical and administrative tasks, where the company says access limits and human approvals can be set around an agent’s actions.

GPT-6 Astra faces clinical and administrative tests

OpenAI evaluated GPT-6 Astra on EHR inbox tasks shaped by physician input. Reviewers examined the agent’s actions together with the notes, orders and messages it produced. In the company’s results, 99% of 325 reviewed tasks were rated clinically safe, and physicians called its work “very helpful” or “extremely helpful” in 93% of 646 assessments. The safety assessment considered whether inaccuracies, omissions or actions could plausibly harm a patient.

Karan Singhal, OpenAI’s head of Health AI and Artificial General Intelligence, said researchers found no specific, plausible risk of patient harm among the tasks rated unsafe. Singhal said the model might, in some cases, take a step without the necessary sign-off or authorisation. That failure belongs alongside the safety rating: a task can require an approval even when the resulting clinical content presents no identified risk of harm.

On HealthAdminBench, a benchmark developed by Stanford researchers for administrative workflows, OpenAI reported that Astra earned 94.3% of available rubric points and passed every required check in 60.7% of task attempts. Rubric points award credit across a task’s criteria; passing an attempt requires every mandatory check. The benchmark covers work such as prior authorisations, appeals and medical equipment orders.

More than 20 ChatGPT for Healthcare plugins

OpenAI says the new plugins are available in ChatGPT for Healthcare, including through ChartGPT within an EHR workflow. They include Luma Health for appointment information, Metriport for records from elsewhere and Stedi for insurance eligibility. OpenAI describes uses ranging from reviewing hospital protocols and preparing prescriptions for clinician review to coordinating outreach and examining healthcare costs.

Clinical search has also been updated in ChatGPT for Healthcare and ChatGPT for Clinicians. OpenAI says the feature finds and cites medical evidence, giving clinicians sources they can inspect. In a blind evaluation covering real-world questions across multiple specialties and conversations with several exchanges, physicians rated the updated experience more highly across all eight dimensions assessed in over 1300 reviews. Physicians assessed factual correctness, safety, actionability and the authority of the medical evidence.

OpenAI says organisations handling protected health information must use services eligible under their business associate agreement with the company, configure the required safeguards and data-keeping rules, and carry those requirements through to other systems used in the workflow. OpenAI says content shared with ChatGPT for Healthcare is not used to train its models.

Sources

Topics: Agents, Enterprise adoption, Foundation models, Healthcare

What should we cover? Tell us

The weekly briefing

The AI Affairs weekly briefing: the week's 10 most important AI stories in business, government and markets, every Monday. We'll email you a link to confirm. Every email has a one-click unsubscribe. Privacy policy