Did you try Opus-4.7 on a lower reasoning level? Looks like on `max` it's using far more tokens than the other frontier models.