What it is
A chain of thought is the series of intermediate steps a model writes down while working toward an answer. Instead of jumping straight to a result, the model "thinks out loud": it breaks a problem into parts, works each one, and then combines them.
Think of a math teacher who says "show your work." The final answer matters, but the work is how you catch mistakes.
How it works
A language model produces text one token at a time. Every token it writes becomes part of its own input for the next one. When a model writes out intermediate steps, those steps give it more room to compute, because each step conditions the next. On multi-step problems such as math, logic puzzles and planning, that extra working space tends to produce better answers than a one-line reply.
You can ask for it in a prompt ("work through this step by step"). Many newer systems are also trained to do it on their own. That is the idea behind a reasoning model, which spends extra tokens on a visible or hidden chain of thought before it replies.
Why it matters to you
- Accuracy. For tasks with several steps, a chain of thought usually helps. For simple lookups or short rewrites it only adds cost.
- Cost and speed. The steps are tokens, and tokens are what you pay for and wait on. Longer reasoning means a slower, pricier answer, and it also uses room in the context window.
- Oversight. If a model's reasoning is readable, people can check it for mistakes or for signs the model is pursuing the wrong goal. Researchers call the ability to read and trust that reasoning "monitorability."
A caution
A written chain of thought is not always a faithful record of why a model produced its answer. The text can read as tidy and convincing while leaving out what actually drove the result. Treat it as a useful clue and a way to spot errors, not as proof.
Example
Ask a model, "A shirt costs $40 after a 20% discount. What was the original price?" A direct guess might say $48. With a chain of thought, the model writes: the sale price is 80% of the original, so the original is 40 divided by 0.8, which is $50. Seeing the steps makes it easy for you to catch a wrong turn.
Chain of thought is the core technique behind the reasoning model. Its safety value connects to AI alignment, since readable reasoning is one way to check what a model is trying to do.