What excites me most about these new 4figure/second token models is that you can essentially do multi-shot prompting (+ nudging) and the user doesn't even feel it, potentially fixing some of the weird hallucinatory/non-deterministic behavior we sometimes end up with.