OpenAI Launches Computer-Use Agents for Healthcare, Citing Safety Benchmarks and Early Partner Deployments
OpenAI has announced computer-use agents designed to operate across electronic health records, insurer portals, and other clinical software, positioning them as a way to automate administrative workflows such as prior authorization, referral management, and discharge planning. The company published two blog posts today describing the capabilities, evaluation results, and early deployments with CVS Health, Abridge, Cigna Healthcare, and Ambience.
The announcement centers on "computer-use agents" that click, type, and navigate between applications the way a human user would, rather than relying on APIs. OpenAI says the agents are available now through ChatGPT for Healthcare, ChatGPT for Clinicians, and the OpenAI API, with HIPAA-compliant configurations supported under a Business Associate Agreement.
Claimed Performance and Safety Metrics
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 the model to navigate EHR inboxes and carry out multi-step workflows. Reviewers examined the model's actions alongside the notes, orders, and messages it produced. OpenAI's blog post describes a separate demonstration using synthetic patient data, but does not state what data was used for the clinical safety and helpfulness evaluations.
On HealthAdminBench, an independent benchmark from Stanford researchers measuring administrative workflow performance, OpenAI says GPT‑6 Astra scored 94.3% of available rubric points while GPT‑6 Luna scored 91.8% at 8.5 cents per task, a 98% cost reduction. Both were evaluated at maximum reasoning effort.
These figures come from OpenAI's own evaluation process. The company does not describe how reviewers were selected or what data the evaluation tasks used. HealthAdminBench is described as independent, but the benchmark results are reported by OpenAI without third-party verification in the posts.
Partner Deployments: What Is Actually Running
OpenAI highlights four organizations building with the computer-use capability:
CVS Health says it has a production workflow in which an agent cleared a months-long document-sending backlog over a weekend, and another that handles "tens of thousands of consults each day." CVS Health states it uses validation checks and human oversight with screenshots and action records, and that production data is not used to train foundation models. The company is open-sourcing an "Agent Kit" built with OpenAI's Agents SDK.
Abridge is developing agents for inpatient discharge planning, combining clinical conversation data with EHR information to coordinate medication access and home-health referrals. The company says clinicians stay focused on the patient and are brought in for decisions requiring clinical judgment.
Cigna Healthcare reports that case managers previously navigated more than 10 separate systems to gather member context. Computer use now helps bring that information together in seconds. A Cigna executive said the deployment moved "from concept to deployment in under eight weeks."
Ambience built additional governance controls around computer use, including step-by-step audit logs, configurable action permissions, and a behavior where the agent pauses for clinician approval when records are ambiguous. In a demonstration, an agent prepares a prior-authorization request on an insurer website and pauses for clinician approval before submitting.
Each partner has a commercial relationship with OpenAI and an incentive to showcase progress. The CVS Health description of "production workflow" and "tens of thousands of consults each day" is the most concrete claim of scale, but no independent metrics on error rates, time saved, or staffing impact are provided.
Plugin Ecosystem Expansion
OpenAI also announced more than 20 new healthcare plugins for ChatGPT for Healthcare, covering clinical care, care coordination, and business operations. Examples include 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. The plugins are available today, subject to workspace administrator enablement.
Access and Governance Model
Organizations must use eligible services covered by their Business Associate Agreement with OpenAI, apply required security and retention settings, and extend protections to connected systems. ChatGPT for Healthcare provides role-based access controls and audit logs. Individual U.S. clinicians can sign up for a free ChatGPT for Clinicians account. Developers access the capability through the OpenAI API, ChatGPT Work, and Codex.
OpenAI states that organization data is not used to train its foundation models.
What Remains Unclear
- The clinical safety and helpfulness evaluations do not specify the data source or how reviewers were selected. Real-world error rates, edge-case behavior, and liability frameworks are not addressed.
- CVS Health's "tens of thousands of consults each day" and weekend backlog clearance are described without independent confirmation or detail on what fraction of total volume they represent.
- No timeline is given for broader availability beyond the current ChatGPT for Healthcare, Clinicians, and API channels.
- Pricing for the computer-use capability at scale is not disclosed; the HealthAdminBench cost figure of 8.5 cents per task applies to GPT‑6 Luna at maximum reasoning effort on a specific benchmark.
- The posts do not discuss integration with specific EHR vendors (Epic, Cerner, etc.) beyond the partner demonstrations.
OpenAI's announcement marks a shift from API-dependent integration to computer-use agents that operate existing software interfaces directly. The claimed benchmarks and partner deployments suggest early traction, but the evidence base consists of self-reported evaluations, a synthetic-data demonstration, and partner statements without independent audit. Healthcare organizations considering adoption will need to assess governance controls, audit requirements, and real-world performance in their own environments.
Sources
- Hermes Agent vs ChatGPT dots: an open-source alternative (hermes-agent.nousresearch.com)
- OpenAI Launches Business Hub and Business Agents in ... (unite.ai)
- ChatGPT (primo.ai)
- ChatGPT Dots: What They Are & How to Use Them (4geeks.com)
- Advancing computer-use agents for healthcare teams | OpenAI (openai.com)
- More ways to move care forward with ChatGPT for Healthcare | OpenAI (openai.com)