OpenAI Opens Computer-Use Agents to Health Teams, Says Astra Was Safe on 99% of Tasks
OpenAI made computer-use agents available to healthcare teams through ChatGPT for Healthcare, ChatGPT for Clinicians and its API, with HIPAA-eligible configurations, approvals and audit logs. In OpenAI's evaluation, reviewers rated 99% of 325 GPT-6 Astra EHR inbox tasks clinically safe, and on Stanford's HealthAdminBench Astra earned 94.3% of rubric points against 91.8% for GPT-6 Luna at 8.5 cents a task. CVS Health says an agent cleared a months-long document backlog over a weekend. The evaluations are OpenAI's own, and OpenAI told Fierce Healthcare the unsafe 1% included actions taken without proper sign-off.

The pitch for computer-use agents in healthcare is that they reach the screens APIs never did: the payer portal, the EHR inbox, the system with no integration. OpenAI says those agents are now available to healthcare teams through ChatGPT for Healthcare, ChatGPT for Clinicians and its API, and backs the launch with safety and cost numbers from its own evaluations. The 99% figure is the headline; the 1% is where the risk sits, and OpenAI says some of it was agents acting without the proper sign-off.
What OpenAI is offering#
Organisations can configure which applications and actions agents may use, add approval steps and record audit logs. To handle protected health information, they must use services covered by a Business Associate Agreement with OpenAI and the required security and retention settings. US clinicians can use it through a free ChatGPT for Clinicians account.
The numbers, and whose they are#
- Clinical inbox tasks: reviewers rated 99% of 325 GPT-6 Astra tasks clinically safe, judged by whether inaccuracies, omissions or actions could plausibly lead to patient harm. Physicians rated the work "very helpful" or "extremely helpful" in 93% of 646 assessments.
- Administrative tasks: on HealthAdminBench, a benchmark from Stanford researchers covering prior authorisations, appeals and equipment orders, OpenAI says Astra earned 94.3% of rubric points and GPT-6 Luna 91.8%, with Luna costing 98% less at 8.5 cents a task. Both ran at maximum reasoning effort.
On the 1%, OpenAI's Karan Singhal told Fierce Healthcare researchers found no case where the model created a specific, plausible risk of patient harm, but that "in some cases, the model might overreach and might take an action without getting the proper sign-off or authorization."
Early deployments#
CVS Health says one production agent cleared a months-long document-sending backlog over a weekend and another handles system steps for tens of thousands of consults a day, with validation checks and human oversight; it is open-sourcing an Agent Kit built on OpenAI's Agents SDK. Cigna Healthcare says computer use helps case managers pull context from more than 10 systems. Abridge and Ambience describe agents still in development or demonstration.
Google made its own computer-use tool generally available in the Gemini API at the end of October; OpenAI's release is aimed at a single regulated sector rather than general developers.
Limits#
- The safety and helpfulness ratings come from OpenAI's evaluation with reviewers it chose; we have not seen the task set or an independent replication.
- HealthAdminBench is independent, but the scores were run and reported by OpenAI. Astra and Luna were both at maximum effort, so the cost comparison applies only at that setting.
- Partner results are self-reported, and some described uses are pilots or demos, not production.