which is largely due to model size

I disagree. I would argue GPT-4 is useful because of the sea of augmentations built up around it.

If it was just raw responses from a single base model (like most local LLMs), with no preprocessing, I believe GPT-4 would be much less impressive.