OpenAI pitches GPT-6 Astra and Luna agents for medical forms and insurance requests

OpenAI wants GPT-6 Astra and GPT-6 Luna to click and type their way through clinic software and insurance portals as digital clerks. According to the company, GPT-6 Luna handles administrative tasks for about 8.5 cents per task. The Decoder reports on a recent OpenAI product announcement that targets the lack of APIs in healthcare. The computer-use agents, integrated into the Agents API in late September, operate existing user interfaces the way a human would. Here, that means moving data between electronic health records and external web portals.
OpenAI's example: when a patient reports back pain, the agent is meant to search relevant prior findings and draft both an MRI order and a reply to the patient. It should also prepare a prior authorization request on the insurer's portal. Clinicians and the responsible administrative staff review these drafts before they are released. The features are available through the API, and in ChatGPT for Healthcare and ChatGPT for Clinicians.
OpenAI backs the claims with two sets of numbers. In an evaluation it describes, GPT-6 Astra worked through tasks from electronic health record inboxes, among them preparing an ibuprofen dosage for a child with a broken arm. Physician reviewers rated 99 percent of the 325 tasks evaluated as clinically safe. In 93 percent of 646 reviews, the reviewers called the prep work very or extremely helpful. For administrative workflows such as prior authorizations, appeals and medical equipment orders, OpenAI points to the independent HealthAdminBench benchmark from Stanford University researchers. GPT-6 Astra scored 94.3 percent of possible points. GPT-6 Luna, built for low API pricing, scored 91.8 percent and, according to OpenAI, cost about 8.5 cents per task, 98 percent less than Astra. Both models ran at maximum reasoning effort on the benchmark.
Partners are named. At CVS Health, the agents already process referrals and specialist consultations across separate systems. According to the announcement, one agent cleared a months-long backlog of outgoing documents in one weekend, while another handles the system steps for tens of thousands of consultations daily. CVS Health published its software architecture as Agent Kit on GitHub. At insurer Cigna Healthcare, the models merge patient information from more than 10 separate systems. Abridge is developing the agents for planned use in clinical discharge planning. Ambience has demonstrated safety mechanisms in which an agent stops when a patient's record is unclear and asks a physician for input. Startups like Assort use the feature in Codex to complete referrals in practice management software whose APIs don't support that final click.
Privacy is the core concern, since the agents work directly in health records. OpenAI says it does not use organizational data to train its base models. ChatGPT for Healthcare offers role-based access controls and audit logs, according to the company, and organizations can configure environments so an agent only reaches the records and actions a task needs. For HIPAA-compliant workflows with protected health information, organizations must sign a Business Associate Agreement (BAA) with OpenAI and use the covered services with the required security and retention settings. Healthcare facilities are responsible for proper use, but OpenAI, as a contracting party, has its own HIPAA obligations. Verified medical professionals in the US who use ChatGPT for Clinicians, a free individual plan, can also sign a BAA, though it covers only the features the agreement covers. In the CVS Health deployment, data stays within infrastructure that CVS controls, according to OpenAI, and CVS reviews the agents' work steps through screenshots and action logs.
The announcement does not say where those screenshots are processed or how long they are stored, and it does not address the European GDPR. So far, the offering targets the US market. The Decoder notes that broad adoption likely will not depend solely on the accuracy rates and costs in the announcement. It also depends on how reliably the agents connect to existing systems and whether AI agents can be dependably monitored in practice. Whether a human is required to decide on clinically or legally relevant steps remains an open question.
Key facts
- OpenAI's GPT-6 Astra and GPT-6 Luna operate clinic software and insurance portals through mouse clicks and keystrokes, available via the Agents API, ChatGPT for Healthcare and ChatGPT for Clinicians.
- On Stanford's HealthAdminBench, GPT-6 Astra scored 94.3 percent and GPT-6 Luna 91.8 percent, both at maximum reasoning effort; OpenAI says Luna costs about 8.5 cents per task, 98 percent less than Astra.
- In OpenAI's evaluation of GPT-6 Astra, physician reviewers rated 99 percent of 325 health-record inbox tasks clinically safe, and 93 percent of 646 reviews called the prep work very or extremely helpful.
- CVS Health already uses the agents for referrals and specialist consultations; Cigna Healthcare merges patient data from more than 10 systems; Abridge, Ambience and Assort are also named.
- The offering targets the US market; the announcement does not say where CVS screenshots are processed or how long they are stored, and does not address GDPR.
Why it matters
Healthcare administration runs on portals and record systems that often have no APIs. OpenAI's pitch is that an agent which clicks and types like a human can bridge that gap without waiting for integrations. The announcement pairs that pitch with named deployments at CVS Health and Cigna Healthcare, and with a cheaper model, GPT-6 Luna, that OpenAI says costs about 8.5 cents per administrative task.
Who it affects
Clinic and hospital administrative staff, clinicians who review drafted orders and patient replies, and insurers whose portals receive prior authorization requests. Named partners are CVS Health, Cigna Healthcare, Abridge, Ambience and Assort. The offering currently targets the US market and its data privacy requirements, so organizations outside the US are not the audience yet.
How to use it
The features are available through the Agents API and in ChatGPT for Healthcare and ChatGPT for Clinicians. For HIPAA-compliant workflows with protected health information, an organization must sign a Business Associate Agreement with OpenAI, use the covered services and apply the required security and retention settings. Verified US medical professionals on ChatGPT for Clinicians, a free individual plan, can also sign a BAA, limited to the features it covers. Organizations can configure applications and environments so an agent reaches only the records and actions a task needs. In OpenAI's described workflow, clinicians and responsible administrative staff review drafts before release.
How solid is it
Mostly first-party. The 99 percent clinically-safe and 93 percent helpful figures come from an evaluation described by OpenAI and cover GPT-6 Astra. HealthAdminBench is an independent benchmark from Stanford University researchers, but the 8.5 cents per task and the 98 percent cost gap are OpenAI's account. The CVS backlog and volume figures come from the announcement, with no independent confirmation given. The source reports no errors, incidents or patient harm from the deployments.
Risks and caveats
Health records are sensitive, and the agents work directly in them. OpenAI says it does not train base models on organizational data and offers role-based access controls and audit logs. The announcement does not say where CVS's review screenshots are processed or how long they are stored, and it does not address GDPR. The Decoder says adoption likely will not rest on accuracy and cost alone: it also depends on how reliably the agents connect to existing systems and whether they can be dependably monitored. Whether a human must decide on clinically or legally relevant steps is an open question.
“Another open question is whether a human is required to decide on clinically or legally relevant steps.”
— The Decoder