Models consume more tokens than ever for the same tasks.