The reasoning revolution in large language models has so far run on a single trick: make the model talk to itself. Chain-of-thought prompting, and the reasoning models built on top of it, gets better answers by having the model write out intermediate steps as text before committing to a final response. It works remarkably well, and it is remarkably wasteful. Every step of thought is a full pass through the model to produce a token, and every token costs money, latency and context-window space.