Urdu Pro
Pakistan News, Tech, Sports & Showbiz

OpenAI Puts AI Agents Inside Hospitals: GPT-6 Astra Rated Safe on 99% of 325 Tasks

OpenAI healthcare AI agents have moved out of the chat window and into the hospital. On 16 November, the company announced that computer-use agents are now available for healthcare teams — software that reads screens, clicks buttons, types into forms and switches between applications, doing the multi-step computer work that eats up hours of staff time every day.

The agents are powered by GPT-6 Astra, with a cheaper sibling, GPT-6 Luna, for high-volume administrative work. They run inside the software hospitals already use, and OpenAI says physicians rated GPT-6 Astra clinically safe on 99% of 325 reviewed tasks. CVS Health and Cigna Healthcare are already running the agents in production. The development follows other hospital AI work, including Pakistan’s first AI hospital.

What the agents actually do

The pitch is not diagnosis — it is paperwork. The agents take on referrals, prior authorizations, discharge plans and specialist-consult steps that force staff to retype the same facts across systems never built to talk to each other.

OpenAI’s own example is an MRI prior authorization: the clinical facts live in the electronic health record, but the insurer wants them retyped into a separate payer portal. In the demo, the agent drafts the order, the encounter note and the patient reply for a clinician to review, then prepares the payer submission.

The agents are available through the OpenAI API, ChatGPT for Healthcare and ChatGPT for Clinicians.

The safety numbers — and what they do not prove

The headline figure comes from OpenAI’s own evaluation of GPT-6 Astra on multi-step EHR inbox workflows shaped with physician input. Reviewers judged whether errors in the model’s work could plausibly harm patients. They rated 99% of 325 reviewed tasks safe, and rated the work “very helpful” or “extremely helpful” in 93% of 646 assessments.

On Stanford’s independent HealthAdminBench, which tests administrative workflows such as prior authorizations, appeals and equipment orders, GPT-6 Astra scored 94.3% of the rubric points. GPT-6 Luna scored 91.8% at about 8.5 cents per task — roughly 98% less than Astra — with both models at maximum reasoning effort.

Those numbers deserve a careful read. They are OpenAI’s reported evaluations, not FDA clearance and not measured patient outcomes. A physician rating a task safe in review is not the same as a patient being safe in a live clinic. OpenAI’s head of Health AI, Karan Singhal, told Fierce Healthcare that no task in the group not rated safe created a specific, plausible risk of patient harm — but the full task set has not been published, and no independent replication exists yet. For a fuller breakdown, see the ExplainX analysis and the Subagentic report.

CVS and Cigna are already running them

CVS Health is the furthest along. OpenAI says one production agent cleared a months-long document-sending backlog over a single weekend, while another handles the system steps behind tens of thousands of specialist consults a day. CVS is also open-sourcing Agent Kit, built with the OpenAI Agents SDK — and data stays in CVS-controlled infrastructure, not used to train foundation models.

Cigna Healthcare has moved from concept to production in under eight weeks. Its care managers previously jumped across more than ten separate systems to build a picture of a member; OpenAI says the agents can now aggregate that context in seconds. Abridge and Ambience are piloting the capability too — Abridge on inpatient discharge workflows, and Ambience on prior-authorization prep that stops and asks when a chart is ambiguous.

More than 20 plugins and better clinical search

The second half of the announcement upgrades ChatGPT for Healthcare itself. More than 20 healthcare plugins are now available, covering EHR access, scheduling, payer work and clinical knowledge — partners include Redox, Stedi, Turquoise Health, UpToDate Expert AI, Metriport, Luma Health, TriZetto and Oracle HDI.

Trusted clinical search has also been improved, with better citations: blind physician reviews across more than 1,300 assessments reportedly rose on eight dimensions, including factual correctness and source authoritativeness. The plugins are designed to work together in multi-step prompts — for example, Luma plus TriZetto for appointment confirmation and eligibility, or Metriport plus UpToDate for post-discharge medication monitoring. Voice can now start an MRI follow-up that hands off to computer use, aimed at clinicians who will not type long prompts between rooms. OpenAI is also extending ChatGPT into everyday tasks, including ChatGPT Wallet, which covers AI-agent shopping and purchases.

Privacy guardrails and what comes next

OpenAI repeats the deployment guardrails: role-based access and audit logs, configurable approvals and application allowlists per workflow, and BAA-covered services only when protected health information is involved. Organization data is not used to train foundation models, and human sign-off remains mandatory — the agents draft and prepare, people approve.

What the announcement does not solve is just as important. Liability and licensure when an agent drafts orders remain open questions, benchmarks are rubric scores rather than appeal win rates, and the announcements emphasize US clinician programmes and HIPAA framing, so cross-border deployment is still an open question.

Elsewhere in AI, ByteDance Seedance 2.5 Flash has launched as a faster, cheaper AI video model.

Even with those caveats, the direction is clear: the AI agent is no longer just answering questions about healthcare. It is sitting at the hospital computer, clicking through the work itself — and the next year of production use will show whether the 99% holds up with real patients on the other side of the screen.

OpenAI healthcare AI agents
Written by UrduPro Editorial Team
Articles on UrduPro are researched and reviewed by our editorial team before publication. See our Editorial Policy and Contact pages for how we work and how to send corrections.
UrduPro Editorial Team
UrduPro Correspondent

Articles on UrduPro are researched and reviewed by our editorial team before publication. See our Editorial Policy and Contact pages for how we work and how to send corrections.

Leave a Reply

Your email address will not be published. Required fields are marked *

URDU PRO

Pakistan News, Tech, Sports & Showbiz

Get the top stories daily

Breaking news and updates, straight to your inbox.

© 2026 Urdu Pro. All rights reserved.