>Frontier models are mostly past the point of human ability to discern whether they are actually better or worse than predecessors and competitors.

Yeah nah, the models' flaws are pretty obvious when you use them. And as a user, you can absolutely know when a flaw disappears or barrier is cleared.