Quantifying the algorithmic improvement from reasoning models