“Reasoning models are significantly more compute-hungry during inference than non-reasoning models.”