What is a Chain of Thought?

Language models trained or prompted to produce chain-of-thought output effectively “think out loud” — breaking a problem into sub-steps, working through each one, and arriving at an answer that's grounded in the visible reasoning chain rather than pattern-matched in one leap. This measurably improves performance on tasks requiring multi-step logic, arithmetic, or planning, because the model can catch its own errors partway through rather than committing to a wrong answer immediately. Chain of thought also has a practical transparency benefit distinct from its accuracy benefit: a visible reasoning trace lets a human or another system audit how the model arrived at a conclusion, which matters enormously for agentic systems making consequential decisions. It's not free, though — generating the intermediate steps costs additional tokens and latency, so it's applied selectively to problems complex enough to benefit from it rather than universally to every model call.