"...the model may give different advice even when the underlying question is the same."

Isn't this the point of LLMs? If not it would be deterministic and that's not "new" and/or "exciting".