> For scientific questions (ones that should have publicly available data, not ones relying on internal data), I have started going to 2.5 Pro over senior experts on my own team.

Have you asked any of your experts to double check those bot answers to see how it did?