Reasoning models (o1, DeepSeek-R1, Gemini Thinking, Claude with extended thinking) build an internal chain of thought before answering, delivering a leap in multi-step math, code and logic. The 2026 version of this: “thinking” became a toggle — you pick instant reply or deep deliberation.
Mechanism: training on reasoning chains + extra test-time compute. Same model, smarter answers — the test-time-compute law in action.
The consumption call: for 90% of work (summarize, translate, chat) a fast general model; for hard, expensive problems, reasoning mode. Paying for thought makes sense where mistakes are costly.
