The blog post says that their gains come largely from the model trying harder. So, more tokens, more time spent.