I agree with the OP that I can get LLM's to do things now that I wouldn't even attempt a year ago, but I feel it has more to do with my own experience using LLM's (and the surrounding tools) than the actual models themselves.

I use copilot and change models often, and haven't really noticed any major differences between them, except some of the newer ones are very slow.

I generally feel the smaller and faster ones are more useful since they will let me discover problems with my prompt or context faster.

Maybe I'm simply not using LLM's in a way that lets the superiority of newer models reveal itself properly, but there is a huge financial incentive for LLM makers to pretend that their model has game-changing "special sauce" even if it doesn't.