+1. I wish Gemini 2.5/3 Pro's "personality" and long context handling wasn't so erratic, because the medical stuff in there is great. Whatever they did to produce the MedGemma models is clearly built on a strong baseline. I haven't had need to try using MedGemma on x-ray imagery, but I'd be curious to hear results — imagery diagnostics is part of what it's built for.

Opus 4.5 seems good too, though getting dumber. OpenAIs fine tuning is clearly built to toe the professional medical advice line, which can be good and bad.