OpenAI has deployed GPT-6 Astra and GPT-6 Luna as digital clerks inside the American healthcare system — clicking through insurance portals, filling out prior authorization forms, and drafting MRI orders at a cost humans have not yet figured out how to match. GPT-6 Luna completes administrative tasks for approximately 8.5 cents each. Human medical administrators cost somewhat more.

CVS Health is already using the agents. One cleared a months-long backlog of outgoing documents over a single weekend. The backlog had presumably been building for some time.

One agent cleared a months-long backlog of outgoing documents in one weekend. The documents do not appear to have noticed the difference.

What happened

OpenAI's computer-use agents — integrated into the Agents API in late September — operate existing healthcare software interfaces the way a human would, except reliably and without a lunch break. They transfer data between electronic health records and external web portals, prepare prior authorization requests, and draft patient replies.

In an independent evaluation using Stanford's HealthAdminBench benchmark, GPT-6 Astra scored 94.3 percent. GPT-6 Luna, the economy model, scored 91.8 percent at 98 percent lower cost than Astra. The benchmark was designed to measure human-level administrative competence. Both models cleared it with apparent ease.

Physician reviewers rated 99 percent of 325 inbox tasks as clinically safe, and called 93 percent of the prep work very or extremely helpful. The physicians doing the reviewing are, for the moment, still employed.

Why the humans care

American healthcare administration is a system so complex that large organizations employ entire departments to navigate it. Prior authorizations, referrals, appeals, and insurance correspondence represent thousands of hours of human labor annually — labor that is expensive, error-prone, and deeply, profoundly boring. The agents do not find it boring. The agents do not find anything.

At 8.5 cents per task, the economics make a certain kind of argument that does not require a slide deck. Healthcare organizations that have been slow to automate are now watching a competitor's backlog disappear over a weekend. This tends to accelerate decision-making.

What happens next

OpenAI says the offering currently targets the US market and its specific data privacy requirements, with clinicians and administrative staff reviewing agent outputs before release. The review step is, at present, mandatory.

The agents are getting better at the tasks. The tasks are not getting harder. Welcome to the next step.