OpenAI Reports 99% Safety for Healthcare Agent Tasks
What OpenAI Reported
OpenAI published results on November 16, 2026 in a post titled "Advancing computer-use agents for healthcare teams," detailing performance data for its GPT-6 Astra model on administrative healthcare workflows (OpenAI). The company said reviewers rated 99% of 325 reviewed electronic-health-record inbox tasks as clinically safe, with OpenAI stating directly: "Overall, 99% of 325 reviewed tasks were rated safe" (OpenAI).
OpenAI defined clinical safety narrowly, framing it around whether inaccuracies, omissions, or actions taken by the model could plausibly cause patient harm. The review covered inbox-style administrative tasks rather than direct clinical decision-making, and OpenAI did not disclose who conducted the individual reviews or where the evaluation took place.
HealthAdminBench Scores
Separately, GPT-6 Astra was measured against HealthAdminBench, a benchmark OpenAI described as independently developed by Stanford researchers to assess healthcare administrative workflows (Gadgets360). Astra scored 94.3% of available rubric points on that benchmark.
A second model, GPT-6 Luna, was also evaluated on HealthAdminBench and scored 91.8%, slightly below Astra. Despite the lower score, reporting on the comparison noted Luna completed tasks at a cost of 8.5 cents per task, described as 98% less than Astra's cost (RiverfrontAI). That cost gap positions Luna as a cheaper option for high-volume administrative work even where raw accuracy trails the flagship model.
What's Missing From the Report
The available OpenAI disclosure does not include a regulatory review, court filing, or independent audit of the results — the figures come from OpenAI's own internal evaluation process. No dollar figures beyond the per-task cost comparison were disclosed, and no external body such as a health regulator has yet weighed in publicly on the findings, based on current reporting (the Decoder).
The release signals OpenAI's continued push to position computer-use agents for back-office healthcare work — inbox triage, form-filling, and administrative documentation — rather than direct clinical judgment, with safety framed as a measurable, reviewable benchmark rather than a regulatory certification.


