"Benchmaxxing" is the right word for what's happening. I've been evaluating these models on real production tasks, and the gap between leaderboard scores and actual reliability is widening, not shrinking. Trust but verify, always.