I can't help but feel like some sort of hybrid approach: use GPT5 for the strategy, then a more direct ML model for actually executing the strategy might work better than trying to use reasoning directly for input control, would work better than trying to reason your way through driving.

That's what the article describes, yes.