235B-A22B-2507: significantly less token usage than surrounding thinking models.