And to be more specific, the fact that reasoning models have outputs orders of magnitude longer than other models.

It writes more text so it spends more compute.