
We benchmarked Gemini 3.7 Flash on two medical evals: HealthBench Professional (clinician-authored cases) and MedAgentBench (agentic EHR tasks) 🏥 A clear, consistent step up from Gemini 3.6 Flash on both, strong agentic results with a 🥉 third-place finish on MedAgentBench and landing mid-pack on HealthBench Professional. HealthBench Pro: the gain is real, not noise, Gemini 3.7 Flash's worst run beat 3.6's best run across three independent runs, on both raw and length-adjusted scores. Answer style unchanged, so it's pure quality improvement. MedAgentBench: ~93% pass@1 (avg of 3× 300-task runs), up ~2.7 pts from 3.6 Flash, with gains concentrated in harder multi-step tasks like recent-lab retrieval and potassium-replacement ordering. 📊 Leaderboards: medicalsphere.ai/benchmarks



















