Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableOpenAI Blog→ original

OpenAI agent fills out insurance paperwork: Luna scores 91.8% for 8.5 cents

To get an insurer's approval for an MRI, a clinic employee has to copy data by hand from the medical record into a portal, and those portals often have no API. OpenAI trained agents to do this by clicking through the interfaces, with a human checking the result. On Stanford's HealthAdminBench, GPT-6 Luna scored 91.8% at 8.5 cents per task, while Astra scored 94.3%, but Luna is 98% cheaper. For high-volume routine work, the smaller model with human review is the better deal.

AI-processed from OpenAI Blog; edited by Hamidun News
OpenAI agent fills out insurance paperwork: Luna scores 91.8% for 8.5 cents
Source: OpenAI Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceOpenAI Blog ↗
Editor-in-Chief
CVSan agent cleared a months-long document submission backlog over a weekend
Key metric #2Physicians rated 99% of 325 reviewed EHR tasks as safe
Cignagathering data from 10+ systems for case managers now takes seconds

For a patient to get an MRI, the doctor's decision is not enough: the insurer has to approve the scan in advance. The necessary data sits in the electronic health record (EHR), and a clinic employee retypes it by hand into a separate insurer portal. These systems often have no API, so ordinary code cannot automate the job.

OpenAI proposes handing this work to agents that operate a computer the way a human does: they click, type, and switch between windows. In HealthAdminBench, an independent test from Stanford researchers that covers typical administrative tasks at clinics, GPT-6 Luna scored 91.8% of available points at a cost of 8.

5 cents per task. The larger GPT-6 Astra scored 94.3%, but Luna is 98% cheaper, roughly 50 times.

If you are building automation for a regulated industry, this is an example of how to choose a model and keep an agent under control.

What the MRI example looks like

OpenAI built a demo on synthetic data. A patient writes to the doctor about lower back pain. The agent opens the EHR and reads the visit notes, physical therapy history, and previous imaging. It then drafts an MRI referral, a visit note, and a reply to the patient, and the doctor reviews all of it. The agent uses the same data to prepare a prior authorization request in the insurer's portal. An authorization specialist does the final check.

The point is that the agent does not need an API. Many insurer portals and legacy clinical systems offer no integrations, and that is exactly where the most manual time goes.

What the tests showed

The clinical side was tested on tasks drawn from clinicians' EHR inboxes. Physicians helped build them. The tasks require understanding a patient's history and not getting lost in complex interfaces. Experts assessed whether the work was useful and whether the agent's errors, omissions, or actions could harm a patient. Of 325 reviewed tasks, 99% were rated safe. Across 646 usefulness ratings, physicians chose "very useful" or "extremely useful" 93% of the time.

For administrative tasks (prior authorizations, appeals, durable medical equipment orders), the team used HealthAdminBench. Astra scored 94.3% of available points and Luna 91.8%. Both models were tested at the maximum reasoning level. The quality gap is about 2.5 percentage points, while the price gap is 98%. For a large stream of repetitive requests, that is a strong argument for the smaller model.

What partners are already doing

CVS Health uses agents to submit documents, process referrals, and support specialist consultations. In one workflow, an agent cleared a backlog of document submissions that had built up over months, in a single weekend. Another handles the system steps for tens of thousands of consultations a day. To keep the work verifiable, CVS saves screenshots and action records, adds verification steps, and keeps a human in the loop. The company is also open-sourcing its Agent Kit, built on the OpenAI Agents SDK: a ready-made template for similar workflows.

At Cigna, case managers used to have to move through more than ten clinical and administrative systems to assemble context on a customer. Now an agent pulls coverage details, providers, visits, and past inquiries in seconds. Going from idea to launch took less than eight weeks.

Abridge is preparing agents for inpatient care. A physician asks for help removing barriers to discharge, and the agents work in parallel on medication access and home care arrangements, combining interface work with calls and messages. Ambience emphasized controllability. If the record contains an ambiguity, the agent stops and asks the physician, and submits a request only after confirmation. The clinic sets which actions are allowed, can pause the agent at any moment, and gets a step-by-step audit log.

Limitations and practical takeaways

The safety and usefulness figures come from OpenAI's own evaluation, and the demo is built on synthetic data. Even 99% safe tasks means that about three of 325 were questionable, and in medicine a remainder like that cannot be ignored.

The capabilities are available through the OpenAI API, ChatGPT for Healthcare, and ChatGPT for Clinicians. The products include role-based access, approval configuration, and audit logs. Organizations' data is not used to train the base models. To work with protected health data under HIPAA (the US health privacy law), organizations must use services covered by a BAA with OpenAI and enable the required security and retention settings.

For developers and product teams, the key point is not the button-clicking itself but the control layer around it: stopping when something is unclear, an action log, access restrictions, and human review. This pattern fits any regulated field where there is no API and a lot of routine work. The cheap Luna shows that for high-volume tasks it is often better to pick the smaller model and spend the savings on checks.

Story

Key Questions Answered

What the MRI example looks like?

OpenAI built a demo on synthetic data. A patient writes to the doctor about lower back pain. The agent opens the EHR and reads the visit notes, physical therapy history, and previous imaging. It then drafts an MRI referral, a visit note, and a reply to the patient, and the doctor reviews all of it. The agent uses the same data to prepare a prior authorization request in the insurer's portal. An authorization specialist does the final check. The point is that the agent does not need an API.

What the tests showed?

The clinical side was tested on tasks drawn from clinicians' EHR inboxes. Physicians helped build them. The tasks require understanding a patient's history and not getting lost in complex interfaces. Experts assessed whether the work was useful and whether the agent's errors, omissions, or actions could harm a patient. Of 325 reviewed tasks, 99% were rated safe. Across 646 usefulness ratings, physicians chose "very useful" or "extremely useful" 93% of the time.

What partners are already doing?

CVS Health uses agents to submit documents, process referrals, and support specialist consultations. In one workflow, an agent cleared a backlog of document submissions that had built up over months, in a single weekend. Another handles the system steps for tens of thousands of consultations a day. To keep the work verifiable, CVS saves screenshots and action records, adds verification steps, and keeps a human in the loop.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…