Reasoning Models vs CoT Prompts
Extended-thinking models do CoT natively — what that retires, what it doesn't, and how to split work between fast and thinking modes.
By 2026 every frontier lab ships models that reason before answering: Claude with extended thinking, OpenAI's o-series lineage and GPT thinking tiers, Gemini's thinking models. These aren't standard models with a step-by-step habit bolted on — they're trained with reinforcement learning to spend internal 'thinking tokens' exploring a problem, checking work, and backtracking before committing to an answer. You typically control the behavior with a budget or effort setting rather than a magic phrase, you're billed for the thinking tokens, and what you see is often a summary of the reasoning rather than the raw trace. The reasoning moved from your prompt into the model.
That retires part of Module 2. 'Think step by step' is redundant when thinking is native — provider docs for reasoning models mostly advise dropping manual CoT instructions — and prescriptive reasoning scripts ('first do X, then Y, then Z') can even underperform by fighting the model's trained strategies. Your lever moves from eliciting reasoning to budgeting it: minimal effort for easy calls, generous budgets for genuinely hard problems. And don't parse the visible thinking — it's summarized, not contractual, and can change between model versions. Output contracts belong on the final answer, exactly as before.
What prompting still owns: everything reasoning can't infer. A crisp problem statement — thinking harder about a vague goal produces elaborate answers to the wrong question. Context and constraints — no amount of reasoning derives facts about your business it was never given. Domain policies — 'flag anything touching PII' isn't deducible from first principles. Output contracts, tone, audience. Decomposition survives too, one level up: reasoning models handle the sub-steps you used to spell out, but sequencing a genuinely huge task into stages with verification between them is still your job. Priming, grounding, format — the whole non-reasoning toolkit — transfers untouched.
The practical decision is routing. Thinking tokens cost real money and real seconds, so 'reasoning model for everything' is as wrong as 'never'. The pattern that's become standard: a fast model or minimal thinking budget for classification, extraction, formatting, and easy chat; escalation to extended thinking for multi-step analysis, debugging, math, and anything where a wrong answer is expensive. Some stacks route automatically, with a cheap model triaging difficulty first. Your CoT skills didn't expire — they became the judgment for when to buy thinking, how much of it, and what to feed it.