OpenAI's GPT-6 agents now do medical paperwork for 8.5 cents a task
OpenAI's GPT-6 Luna handles medical admin tasks for about 8.5 cents each, and CVS Health already runs agents on tens of thousands of consultations daily. The offering targets the US only.
Why it matters
- GPT-6 Luna completes medical administrative tasks for about 8.5 cents per task, 98 percent less than GPT-6 Astra, per OpenAI.
- Physician reviewers rated 99 percent of 325 evaluated tasks by GPT-6 Astra as clinically safe.
- GPT-6 Astra scored 94.3 percent on Stanford's HealthAdminBench; Luna scored 91.8 percent.
- At CVS Health, one agent clears system steps for tens of thousands of consultations daily; another cleared a months-long document backlog in one weekend.
- The agents were integrated into OpenAI's Agents API in late September; the offering targets the US market and does not address GDPR.
OpenAI says its GPT-6 Luna agent completes medical administrative tasks for about 8.5 cents per task — 98 percent less than its sibling model GPT-6 Astra — as the company pushes AI agents into US healthcare paperwork. The agents don't connect through official interfaces. They operate clinic programs and insurance portals the way a human would, with mouse clicks and keystrokes.
The move matters because healthcare is one of the largest and most stubborn administrative cost centers in the US economy, and much of its software lacks the APIs that would normally let automation in. OpenAI is betting that agents driving existing user interfaces can bypass that bottleneck entirely.
What exactly are the agents doing?
OpenAI integrated the computer-use agents into its Agents API in late September. The models work directly in electronic health records and external web portals, transferring data between them.
The company describes a concrete scenario: when a patient reports back pain, the agent searches relevant prior findings and drafts an MRI order plus a reply to the patient. It also prepares a prior authorization request on the insurer's portal. Clinicians and administrative staff review these drafts before anything goes out — humans stay in the approval loop for now.
The features are available through the API and inside ChatGPT for Healthcare and ChatGPT for Clinicians. The takeaway: OpenAI is productizing agent-driven paperwork, not shipping a research demo.
How well do the models perform?
OpenAI's own evaluation gives the strongest numbers in the announcement. GPT-6 Astra worked through tasks from electronic health record inboxes, including preparing an ibuprofen dosage for a child with a broken arm. Physician reviewers rated 99 percent of the 325 evaluated tasks as clinically safe. In 93 percent of 646 reviews, reviewers called the prep work very or extremely helpful.
For administrative workflows — prior authorizations, appeals, medical equipment orders — OpenAI points to HealthAdminBench, an independent benchmark from Stanford University researchers:
- GPT-6 Astra scored 94.3 percent of possible points.
- GPT-6 Luna scored 91.8 percent.
- Both models ran at maximum reasoning effort on the benchmark.
- Luna's per-task cost of roughly 8.5 cents undercuts Astra by 98 percent, according to OpenAI.
The gap between the two models on the benchmark is small; the price gap is enormous. That economics is the core of the pitch to healthcare operators processing paperwork at volume.
Who is already using them?
Several major US healthcare organizations are testing or deploying the agents, and CVS Health is the furthest along. At CVS, 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 a single weekend. Another now handles the system steps for tens of thousands of consultations daily. CVS Health published its software architecture as Agent Kit on GitHub.
Other deployments named by OpenAI:
- Cigna Healthcare: the insurer uses the models to merge patient information from more than 10 separate systems.
- Abridge: developing the agents for planned use in clinical discharge planning.
- Ambience: has demonstrated safety mechanisms where an agent stops when a patient's record is unclear and asks a physician for input.
- Assort: the startup uses the feature in Codex to complete referrals in practice management software whose APIs don't support that final click.
That last case is telling. The agents' main value proposition is filling exactly the gaps that official integrations leave open.
What about data privacy and HIPAA?
Because the agents work directly inside electronic health records, protecting sensitive health data becomes the central risk. OpenAI says it does not use organizational data to train its base models. ChatGPT for Healthcare offers role-based access controls and audit logs, and organizations can configure applications and environments so an agent accesses only the records and actions a given task requires.
For HIPAA-compliant workflows involving protected health information, organizations must sign a Business Associate Agreement (BAA) with OpenAI, use the covered services, and apply the required security and retention settings. The healthcare facilities themselves remain responsible for proper use — but OpenAI, as a contracting party, carries its own obligations under HIPAA as well.
Verified medical professionals in the US using ChatGPT for Clinicians, a free individual plan, can also sign a BAA, though it covers only the features the agreement includes.
Does this reach Europe?
Not yet, and possibly not soon. 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.
It also does not address the European General Data Protection Regulation at all. The offering currently targets the US market and its data privacy requirements. For now, European providers are watching from the sidelines.
Why this is about more than accuracy scores
OpenAI frames the technology primarily as a tool against administrative friction in a highly fragmented US healthcare system. The fragmentation is real: referrals, authorizations and records routinely live in a dozen disconnected systems, which is precisely why agents that click through existing interfaces have an opening.
But broad adoption likely won't hinge on the accuracy rates and costs in the company's announcement alone. Also decisive, per the source's assessment, is how reliably the agents connect to existing systems and whether AI agents can be dependably monitored in day-to-day practice. Another unresolved question is whether a human must decide on clinically or legally relevant steps — a regulatory and liability issue the announcement leaves open.
The 8.5-cent task and the 99 percent safety rating are strong opening arguments. Whether they translate into trust inside hospital workflows is the test CVS Health, Cigna and the rest are now running in production.
Original: openai.com
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
749 articles
Related articles
- OpenAI Says 99% of Reviewed GPT-6 Astra EHR Tasks Were Clinically Safe
- OpenAI Launches 'OpenAI for Healthcare' With HIPAA Support
- OpenEvidence Gives Clinicians Free Access Intelligence Suite
- OpenAI launches ChatGPT agent, folding Operator into ChatGPT
- OpenAI adds over 20 healthcare plugins and upgrades clinical search