And this varies widely between models based on how they are told to reason.