The abstract is read at the publisher; the summary is JClub's.
Gemini 3 Pro and Claude Opus 4.5 completed far more simulated ES/GS prior authorization tasks than Gemini 3 Flash, but all models showed systematic bias toward submitting forms.
- Why it matters: Prior authorization for exome or genome sequencing delays rare disease diagnosis, and LLM browser agents’ clinical reliability is not well characterized.
- What they did: In a sandbox with a simulated payer portal and a synthetic EHR of 836 profiles (compliant and deficient), Gemini 3 Pro, Gemini 3 Flash, and Claude Opus 4.5 were scored on completion, form accuracy, and appropriate withholding, with replication on 100 profiles.
- The result: Gemini 3 Pro (95.5%) and Claude Opus 4.5 (93.7%) completed more tasks than Gemini 3 Flash (56.1%), while correct withholding for deficient profiles was low in the first run (0%, 17.3%, 5.2%) and varied on replication (28.8–36.8% for Gemini 3 Pro), with only 18.0–48.0% identical outcomes across three runs.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.