To validate autonomous navigation across high-stakes clinical and back-office environments, OpenAI evaluated its frontier models—GPT-6 Astra and GPT-6 Luna—at maximum reasoning effort across both EHR inbox triage and standardized administrative workflows.
In physician-designed evaluations simulating complex EHR inbox tasks, GPT-6 Astra executed multi-step workflows requiring longitudinal chart synthesis and user-interface navigation—such as reviewing pediatric fracture notes and weight logs to draft unsent caregiver ibuprofen dosing instructions, cast care guidance, and separate unsigned chart notes pending clinician sign-off. Across 325 reviewed clinical tasks, independent reviewers rated 99 percent of the model’s actions clinically safe, while physicians rated the agent’s output as “very helpful” or “extremely helpful” in 93 percent of 646 assessments.
On the administrative side, OpenAI evaluated agent performance on HealthAdminBench, an independent benchmark developed by Stanford researchers measuring common revenue cycle and operational workflows including prior authorizations, claim appeals, and durable medical equipment orders. GPT-6 Astra achieved 94.3 percent of available rubric points, while the smaller GPT-6 Luna model captured 91.8 percent of available points at a cost of 8.5 cents per task—representing a 98 percent cost reduction relative to Astra and establishing viable unit economics for high-volume administrative automation.
Leading healthcare payers, retail health enterprises, and ambient clinical intelligence vendors are already deploying OpenAI’s computer-use agents to automate workflows where native EHR or payer APIs remain absent.
At CVS Health, engineering teams deployed computer-use agents across internal clinical systems to automate document transmission, referral processing, and specialist consultation workflows. In one production deployment, an agent cleared a multi-month document-sending backlog over a single weekend, while another agent processes the administrative steps for tens of thousands of daily consults under human oversight and screenshot verification.
Source link







