How far can reasoning models scale?