Products & Tools

OpenAI Says 99% of Reviewed GPT-6 Astra EHR Tasks Were Clinically Safe

OpenAI says 99% of 325 reviewed GPT-6 Astra EHR inbox tasks were clinically safe, as CVS and Cigna put computer-use agents into production. Stanford's HealthAdminBench adds outside evidence.

OpenAI Brings Computer-Use Agents to Healthcare Workflows
OpenAI Brings Computer-Use Agents to Healthcare WorkflowsAI-generated
By Elena Vasquez5 min read

Why it matters

  • OpenAI announced healthcare computer-use agents on November 16, 2026, reporting 99% of 325 reviewed GPT-6 Astra EHR inbox tasks were clinically safe.
  • On Stanford's HealthAdminBench, GPT-6 Astra scored 94.3% of rubric points; GPT-6 Luna scored 91.8% at 8.5 cents per task, 98% cheaper than Astra.
  • CVS Health production agents handle system steps for tens of thousands of specialist consults daily; one cleared a months-long backlog over a weekend.
  • Cigna deployed its agent in under eight weeks to replace navigation across more than 10 separate systems.
  • Physicians rated Astra very or extremely helpful in 93% of 646 assessments.

OpenAI reported on November 16, 2026 that 99% of 325 reviewed GPT-6 Astra tasks in electronic health record inboxes were rated clinically safe, and it shipped healthcare computer-use agents through ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API the same day. The figure anchors the company's push into one of the most regulated and error-intolerant software markets in existence. If agents can safely drive the clunky interfaces that dominate hospital administration, they could reclaim some of the hours clinicians lose to paperwork every day.

The agents don't call APIs. They operate existing software interfaces to complete multistep work, including authorization requests, referrals and discharge planning. In a synthetic-patient demonstration, a message about lower back pain prompted a review of visit notes, physical-therapy records and prior imaging. The agent then drafted an MRI order, an encounter note and a patient reply for clinician review, and prepared an insurer-portal authorization request for an authorization specialist to review.

That demo matters because it mirrors the actual workflow: the agent drafts, humans approve. Nothing submits itself in the demonstration scenario.

How Safe, According to Whom?

The safety assessment examined whether inaccuracies, omissions or actions in the model's work could plausibly lead to patient harm. Reviewers assessed Astra's interface actions alongside the notes, orders and patient messages it produced. OpenAI also reports that physicians selected the two highest helpfulness categories — very helpful or extremely helpful — in 93% of 646 assessments on a five-point scale.

The numbers come from OpenAI itself, which is worth keeping in mind. Independent evidence comes from HealthAdminBench, a benchmark from Stanford researchers containing 135 tasks and 1,698 subtasks spanning prior authorization, durable medical equipment orders and appeals.

On that benchmark, OpenAI reports:

  • GPT-6 Astra earned 94.3% of available rubric points, running at maximum reasoning effort.
  • GPT-6 Luna earned 91.8%.
  • Luna's cost of 8.5 cents per task was 98% lower than Astra's.

The benchmark's tasks force agents to observe live web portals through screenshots, accessibility trees or both. Agents then navigate forms, review clinical documentation, check coverage criteria and submit decisions through standard browser interactions. In other words, the models solve the same problem human staff do — portals that were never designed for automation.

What Is Already Running in Production?

CVS Health is the furthest along. OpenAI reports that one CVS Health production agent cleared a months-long document-sending backlog over a weekend. Another handles system steps for tens of thousands of specialist consults each day. These agents navigate existing applications to complete steps that application programming interfaces do not fully support.

CVS Health runs the agents inside guardrails:

  • Validation checks and human oversight verify actions and application state.
  • Screenshots and action records document what happened.
  • Data stays within enterprise-controlled infrastructure, and production data is not used to train foundation models.

CVS Health also open-sourced its CUA Agent Kit under an Apache-2.0 license. The kit uses an OpenAI computer-use agent to discover browser or Windows workflows, then converts successful runs into inspectable Playwright or pywinauto code that runs without model calls. That architecture removes repeated model reasoning from stable workflows. Codex-assisted diagnosis and updates remain subject to engineer review, testing and verification.

The approach is a quiet economics argument: pay for frontier-model reasoning once, then compile it into deterministic scripts.

What Did Cigna Build?

Cigna Healthcare reports that its case managers previously navigated more than 10 separate clinical and administrative systems to gather benefits and coverage, provider, appointment and prior-interaction context. Its deployment assembles that information for member outreach, and it moved from concept to deployment in under eight weeks. The work is part of The Cigna Group's broader collaboration with OpenAI, and it's intended to give case managers more time to support members around care plans established by treating providers.

An eight-week path from concept to deployment is fast for healthcare IT, where integration projects routinely run for quarters. The reason it's possible is that the agents don't require the underlying systems to change.

Where Do Humans Stay in the Loop?

Abridge is developing a cross-system assistant that coordinates discharge work across medication access and home health, with concurrent agents and escalation for clinical judgment. Its own announcement says the experience is live in early environments with a few select partners, and that oversight must show what information was used, whether each step completed, when the assistant escalated and whether a payer received a submission.

Ambience's demonstration applies similar oversight to prior authorization. The agent prepares a request from chart evidence, stops to ask a clinician when the record is ambiguous, and waits for approval before submitting. Health systems can restrict permitted actions, pause the agent and inspect a step-by-step audit log.

The disclosure obligations are specific enough to be useful. A payer either received a submission or it didn't, and the audit trail has to prove which.

What Are the Security Caveats?

CVS Health's own repository warns that run artifacts are not sanitized exports. Traces can contain credentials, cookies, session tokens and network data, while screenshots capture visible information. Credential replacement and redaction markers are best-effort protections rather than comprehensive redaction. Anyone deploying the kit inherits that risk.

On the platform side, ChatGPT for Healthcare provides role-based access controls and audit logs. Organizations can add approval steps and limit agents to the applications, records and actions needed for each task. To support HIPAA-compliant workflows involving protected health information, organizations must use eligible services covered by an OpenAI Business Associate Agreement and the required security and retention settings. OpenAI states that organization data is not used to train its foundation models.

Why the Rollout's Shape Matters

The contrast between CVS Health's production administrative agents and Abridge's early care-coordination environments defines the rollout's central tension. Computer use is already handling high-volume interface work at scale — tens of thousands of consults a day, backlogs cleared over weekends — while broader clinical coordination makes audit trails, access limits and escalation points essential to safe deployment.

The pattern is familiar from other automation waves: administrative drudgery goes first, judgment stays with humans last. How quickly Abridge-style coordination moves beyond a few select partners will depend on whether the oversight mechanisms — audit trails, escalation points, action restrictions — keep proving reliable enough for regulators and health systems to trust.

Source: unite.ai

Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.

708 articles

Related articles

  1. OpenAI's GPT-6 agents now do medical paperwork for 8.5 cents a task
  2. OpenAI Makes ChatGPT for Clinicians Free for Verified US Providers
  3. OpenAI Cancels GPT-6.1 Astra Release Over Deceptive Behavior in Testing
  4. OpenAI Blocks GPT-6.1 Astra Release Over Deceptive Behavior
  5. OpenAI adds over 20 healthcare plugins and upgrades clinical search

« Previous articleNext article »