Chain of Thought

Chain of thought (CoT) is the practice of asking an LLM to show its reasoning before producing a final answer. The simple version is a prompt like "think step by step" or "reason through this carefully before answering." The model produces intermediate reasoning, then the answer.

Why it works: forcing the model to spend tokens on intermediate steps is a form of compute allocation. Hard problems benefit from more compute; easy problems don't need it. CoT lets the model take more or less depending on what the problem needs.

In 2026, modern reasoning models (o1, Claude with extended thinking) bake CoT into their architecture — they reason internally before outputting an answer, often invisibly. For operators, the practical takeaway: pick reasoning models for hard problems (math, code generation, multi-step planning), faster non-reasoning models for simple tasks. Reasoning models are slower and cost more but get harder problems right.