OpenAI describes healthcare agents that click, type, and move among existing clinical and administrative applications to complete multi-step work. In a demonstration using synthetic patient data, an agent drafts clinical materials for a clinician to review and prepares an insurer authorization request for a specialist. The company announced the evaluations and workflows on November 16, 2026, in its overview of computer-use agents for healthcare teams.

What do the agents do in a clinical workflow?

Computer use means navigating software through its existing interface: an agent observes screens or other interface information, then interacts with the application using actions such as clicking and typing. That lets it move between systems even when a workflow involves several steps.

OpenAI’s MRI authorization demonstration starts with synthetic patient records. The agent reviews visit notes, physical-therapy records, and prior imaging, then drafts an MRI order, an encounter note, and a patient reply for a clinician to review. It also prepares an authorization request for an authorization specialist. The example shows a handoff: the agent prepares work, while people review the clinical drafts and the insurer request.

What do the reported evaluations measure?

OpenAI reports that 99% of 325 reviewed GPT-6 Astra clinical tasks were rated safe. Reviewers considered whether inaccuracies, omissions, or actions could plausibly cause patient harm. In a separate set of 646 physician assessments, 93% rated the model’s work “very helpful” or “extremely helpful” on a five-point scale.

Those figures describe reviewers’ judgments of tasks and outputs. They are not patient-outcome measurements. For an administrative benchmark, OpenAI reports these scores for the share of available HealthAdminBench rubric points; it says both models ran at maximum reasoning effort.

ModelAvailable rubric pointsReasoning effort
GPT-6 Astra94.3%Maximum
GPT-6 Luna91.8%Maximum

HealthAdminBench describes agents observing a live web portal through an accessibility tree, screenshots, or both, and interacting with it through standard browser actions. Its scoring includes checks for reaching an exact application state and language-model judgments of clinical accuracy, with reproducibility across runs. That methodology helps explain what the administrative scores cover: interface tasks and their results, rather than patient outcomes. HealthAdminBench’s methodology describes the benchmark’s own approach; OpenAI is the source of the scores above.

OpenAI also reports a benchmark task cost of 8.5 cents for GPT-6 Luna, which it says was 98% lower than GPT-6 Astra’s. That figure is a cost for a benchmark task, not a customer subscription or API rate.

A workflow score is most useful when read alongside the human review step that handles ambiguity before consequential work is submitted.

For broader context on GPT-6 Astra’s computer-use capabilities, see our earlier coverage of the model.

How do healthcare organizations use or develop these agents?

The examples OpenAI describes span production workflows, ongoing development, and demonstrations, so they represent different stages of use.

OpenAI says CVS Health uses agents in production to send documents, process referrals, and support specialist consultations. The company reports that CVS cleared a months-long document-sending backlog over a weekend and uses another workflow for system steps in tens of thousands of consultations each day. CVS’s described checks include screenshots and action records, alongside human oversight.

OpenAI says Cigna Healthcare uses computer use to gather member context from more than 10 clinical and administrative systems for care managers. Cigna Healthcare CIO Madhu Nutakki said the deployment took under eight weeks.

Abridge is developing care-coordination workflows, including a planned inpatient use case for addressing discharge barriers such as medication access and home-health tasks. OpenAI says clinical decisions in that planned workflow would be escalated to a clinician. In an Ambience demonstration, an agent preparing an authorization request pauses and asks a clinician for input when a chart is ambiguous, then waits for approval before submission.

Where do human review and data controls fit?

OpenAI describes configurable permissions and approvals, role-based access controls, audit logs, and limits on which applications, records, or actions an agent can use. In the examples, people review clinical drafts, resolve ambiguity, or approve an authorization request before submission.

For HIPAA-compliant workflows involving protected health information, OpenAI says organizations need eligible services covered by its Business Associate Agreement, along with required security and data-retention settings. OpenAI also says organization data is not used to train its foundation models. For CVS, it says data remains within CVS-controlled infrastructure and production data is not used to train those models.

Which access routes does OpenAI name?

OpenAI lists the API, ChatGPT for Healthcare, and ChatGPT for Clinicians as access routes for the capabilities. It says U.S. clinicians can sign up for a free ChatGPT for Clinicians account and use computer use in ChatGPT Work and Codex.