Chain of thought: when AI reasons step by step
Chain of thought means getting a model to write out the intermediate steps of its reasoning before giving a final answer. It clearly improves results on multi-step problems such as arithmetic, logic and planning.
Data as of October 22, 2026
Key figures
| Level | Intermediate |
|---|---|
| Origin | 2022 |
| Introduced by | Wei et al. (Google) |
The idea in one sentence
Instead of asking for the answer straight away, you let the model “show its working”: it writes out the intermediate steps, then concludes. Like a student laying out a maths problem, it is less likely to skip a step or guess.
A worked example
Question: “A shop sells notebooks at $3 each, and you get one free for every four you buy. How much do ten notebooks cost?”
A model answering in one go might say “$30”, forgetting the offer, or misapply the deal. With chain of thought it reasons: “Pay for four, get five. Ten notebooks is two bundles of five, so I pay for eight. Eight times $3 is $24.” The final answer is now justified and checkable: if a step is wrong, you can see it.
Where it came from
The technique was formalised in the paper “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” (Wei et al., Google, January 2022). The authors showed that giving a model a few examples whose answers are preceded by step-by-step reasoning substantially improves its performance on arithmetic, commonsense and symbolic reasoning tasks. In other words, it started life as a flavour of few-shot prompting.
Why it matters now
Chain of thought has moved from a prompt engineering trick to something built into the models themselves. Reasoning models are trained to produce a long internal reasoning before they answer, often hidden or summarised in the interface. That idea of “thinking for longer” at answer time sits at the heart of test-time compute.
Common myths
- “Detailed reasoning means a correct answer.” A model can produce very convincing steps and still land on a mistake, or even a hallucination. Visible reasoning makes checking easier; it does not replace checking.
- “It helps with every question.” For a translation or a simple factual question the gain is small, and the answer just gets longer and more expensive.
- “It’s the same thing as a reasoning model.” Not quite: chain of thought is a technique; a reasoning model is a model trained to apply it on its own.
Try it yourself
The site’s prompt generator lets you add a step-by-step reasoning request to your prompt, so you can compare answers with and without it.
Related terms
- Prompt engineering · The practice of writing, testing and refining instructions to an AI model so its answers become more reliable and useful.
- Few-shot prompting · A technique where you include a few worked examples of input and expected output in the prompt to show the model what you want.
- Reasoning model · A language model trained to think step by step before answering: slower and pricier, but stronger at maths, code and logic.
- Test-time compute · The computing power a model spends while answering, for instance to think for longer or to compare several candidate solutions.
- Hallucination · An AI answer that sounds confident and plausible but is false or made up: a figure, a quote, a source or a reference.
Try it
Full glossary
Frequently asked questions
Should I still type “think step by step”?
With a reasoning model it is usually unnecessary because the model already does it. With a standard model, asking for detailed reasoning remains an easy way to improve answers on harder problems.
Does the reasoning shown reflect what the model is really “thinking”?
There is no guarantee. The text is a plausible explanation, not a direct window into the model's internal computation. It helps you spot mistakes but does not prove the answer is right.
Is there a cost to chain of thought?
Yes. The model writes more text, so it uses more tokens and takes longer to reply. For a simple question that extra cost buys you nothing.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.