That's an optimistic view. Maybe they really are 10x slower on any task without a LLM.