Comment by MichaelZuo

I imagine the straightforward reason is that the “better” models are in fact significantly smarter in some tangible way, somehow.