OpenAI's Health Agents Click Through EHRs at 8.5 Cents a Task
OpenAI launches computer-use agents for healthcare at 8.5 cents per task, with CVS handling 'tens of thousands of consults daily' — but all evidence is self-reported.
By Nathan Brooks
4 min read

What's News
- OpenAI reports 93% of 646 tasks rated 'very helpful' or 'extremely helpful' and 99% of 325 tasks judged clinically safe
- GPT‑6 Luna scored 91.8% on HealthAdminBench at 8.5 cents per task, a 98% cost reduction
- CVS Health says an agent clears 'tens of thousands of consults each day' and cleared a months-long backlog in one weekend
- Cigna moved its deployment 'from concept to deployment in under eight weeks'
- All benchmarks and partner claims are self-reported without independent verification
By Nathan Brooks
OpenAI says GPT‑6 Luna completed administrative healthcare tasks on Stanford's HealthAdminBench for 8.5 cents each — a 98% cost reduction — as the company launched computer-use agents that navigate electronic health records, insurer portals and clinical software the way a human user would.
The agents, announced in two blog posts, don't rely on APIs. They click, type and move between applications directly. They're available now through ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API, with HIPAA-compliant configurations under a Business Associate Agreement.
How good are the numbers, really?
OpenAI reports that physicians rated 93% of 646 assessed tasks as "very helpful" or "extremely helpful," and that 99% of 325 reviewed tasks were judged clinically safe. The tasks were shaped by physician input and required navigating EHR inboxes and completing multi-step workflows.
On HealthAdminBench, an independent benchmark from Stanford researchers, GPT‑6 Astra scored 94.3% of available rubric points while GPT‑6 Luna hit 91.8% — both evaluated at maximum reasoning effort.
Here's the caveat practitioners should weigh: these are OpenAI's own numbers. The company doesn't describe how reviewers were selected or what data the evaluation tasks used. A separate demonstration relied on synthetic patient data. The benchmark results appear in OpenAI's posts without third-party verification.
What's actually running in production?
Four partners have live or near-live deployments:
- CVS Health says an agent cleared a months-long document-sending backlog over a single weekend, and another workflow handles "tens of thousands of consults each day." CVS uses validation checks and human oversight with screenshots and action records, and says production data doesn't train foundation models. It's open-sourcing an "Agent Kit" built on OpenAI's Agents SDK.
- Abridge is building agents for inpatient discharge planning, combining clinical conversation data with EHR information to coordinate medication access and home-health referrals. Clinicians stay focused on the patient and step in for decisions requiring clinical judgment.
- Cigna Healthcare reports case managers previously navigated more than 10 separate systems to gather member context; computer use now pulls that together in seconds. A Cigna executive said the deployment moved "from concept to deployment in under eight weeks."
- Ambience added governance controls including step-by-step audit logs, configurable action permissions, and an agent that pauses for clinician approval when records are ambiguous — demonstrated with a prior-authorization request that waits for sign-off before submitting on an insurer website.
Remember the incentives here. Every partner has a commercial relationship with OpenAI and a reason to showcase progress. CVS's scale claims are the most concrete, but no independent metrics exist on error rates, time saved or staffing impact.
What does this change for healthcare operators?
The plugin ecosystem expands the practical surface. More than 20 new healthcare plugins for ChatGPT for Healthcare cover clinical care, care coordination and business operations — Luma Health for appointment confirmation, TriZetto for insurance eligibility, Metriport for discharge summaries, UpToDate for clinical guidance, Assort Health for referral closure, and Ferry Health for scheduling follow-ups. They're live today, subject to workspace administrator enablement.
On governance: organizations must use services covered by their Business Associate Agreement, apply required security and retention settings, and extend protections to connected systems. ChatGPT for Healthcare offers role-based access controls and audit logs. Individual U.S. clinicians can sign up for a free ChatGPT for Clinicians account, and developers get access through the OpenAI API, ChatGPT Work and Codex. OpenAI states organization data isn't used to train its foundation models.
What remains unclear
The open questions are material. No data source or reviewer-selection method is disclosed for the safety evaluations. Real-world error rates, edge-case behavior and liability frameworks go unaddressed. No pricing for computer use at scale is published — the 8.5-cent figure applies only to GPT‑6 Luna at maximum reasoning effort on one benchmark. There's no timeline for broader availability, and the posts don't discuss integration with specific EHR vendors like Epic or Cerner beyond the partner demos.
The strategic shift is the story: API-dependent integration is giving way to agents that operate existing software interfaces directly. For healthcare organizations, the practical task now is assessing governance controls, audit requirements and real-world performance in their own environments — not taking the benchmark numbers at face value.
Source: epyhia.com
More from Nathan Brooks
Show full bio
News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.
2,231 articles