This assumes only the LLM matters. I don't think it does. I think GPT3.5 is plenty, if not overkill.

I can't imagine I am alone in my thought. I've even seen some other experiments in the same line of reasoning.

Meta seems to be thinking the same thing, or at least closer to it than the majority.