chain of thought (CoT)

advancedwidespread

Chain of thought is the method where a language model describes step-by-step how it reaches the solution before giving the final answer.

In plain terms

Remember when the math teacher asked you not just to write down the final result, but also the derivation? Chain of thought is the same for language models: the model does not immediately give the answer, but first breaks the task into sub-steps and goes through them one by one.

This is useful because the language model generates the text word by word. If it writes down intermediate steps first, it can rely on them in the continuation. Thus, it makes noticeably fewer mistakes in multi-step calculations, logical tasks, and planning.

At first, you had to request this, for example with the phrase 'think step by step.' Today's so-called thinking models do it on their own: they 'think' before answering, and you often see only the summary of this process.

The described reasoning is not a window into the model's 'mind'. It is text the model generates just like anything else, so it is a useful hint but not proof.

How it works

The model receives the task and, either at your request or due to its built-in operation, does not start with the final answer.
It breaks the problem into parts: what we know, what we are looking for, and in what order to proceed.
It lists intermediate steps in sequence (partial calculation, reasoning, verification), using previously described steps for each new step.
At the end of the steps, it summarizes the result and gives the final answer.
With thinking models, this process takes place in a separate phase that is hidden or shown only as a summary, and it costs more time and compute.

An everyday example

You ask the chatbot whether it is worth prepaying your mortgage. Before the answer, the text 'Thinking...' appears, then the model goes through the remaining principal, interest, and fees, and only then states the amount. This reasoning is the chain of thought.

Who is affected

everyoneIf you use a chatbot, this method stands behind the 'thinking' mode, and the reasoning makes it easier to spot where the answer went wrong.

parent, student, teacherStep-by-step reasoning is good for learning because you see not only the solution but also the path leading to it. → Education & Family

developerAI agents use chain-of-thought to plan which steps and which tools they will use to solve a task. → Code

entrepreneur, executiveThe thinking mode is more accurate but slower and more expensive, so it is worth deciding which tasks justify its use.

A common misconception

MythThe described reasoning process shows exactly how the model arrived at the answer.

In realityChain-of-thought is also generated text that does not necessarily match the model's actual internal calculations. Research shows that models sometimes fabricate convincing reasoning for answers generated for other reasons.

Limits and risks

Plausible reasoning does not guarantee a correct result: a model can confidently derive a wrong answer from a flawed step.
Longer reasoning takes more time and costs more. For simple questions, it barely improves the result and sometimes worsens it, as the model overcomplicates the answer.
Many providers show only a summary, not the full reasoning, so you cannot see everything.
In important decisions (money, health, law), verify the reasoning, not just accept it because it sounds logical.

Similar, but not the same

reasoning model – Chain-of-thought is the method itself, while a reasoning model is a language model specifically trained to apply it even without being asked.
explainable AI – Explainable AI attempts to uncover the basis of a model's decisions using external testing methods. Chain-of-thought, by contrast, is merely the model's own textual account, which no one has verified.
prompt chain – With a prompt chain, you break the task into sequential requests; with a chain of thought, the model proceeds step-by-step within a single response.

Related concepts

Broader

Prompting, large language model, machine inference

Narrower

zero-shot chain of thought, gondolatfa, self-consistency

On the edge

reasoning model, AI-agent, prompt, hallucination, explainable AI

A brief history

The method was described in 2022 by Google researchers (Jason Wei and colleagues). They showed that if a model receives examples including step-by-step reasoning, it performs significantly better on calculation and logic tasks. The improvement was particularly striking in very large models. In the same year, Takeshi Kojima and colleagues demonstrated that the phrase "let's think step by step" is sufficient even without examples.

Where things stand

By the end of 2026, chain-of-thought has become an integral part of leading language models. Most major chatbots offer a thinking mode, and AI agents also build their planning on this. A key research question is how faithfully the described reasoning matches the model's actual operation. Another important question is how effectively reading the reasoning process can filter out unwanted model behavior.

Status: 2026 November

In the news

Appears in 2 Agent Age stories; the curve shows weekly occurrence in the news magazine.

AI visibility market grows: Coty–Brandlight, Marketbridge acquisition, Joveo and GEO start-upsOct 26, 2026 · Marketing
Claim period opens for Apple's $250 million settlement over AI false advertisingSep 21, 2026 · Marketing

Also known as

chain of thought · chain of thought · chain of thought · step-by-step reasoning · chain of thought · chain-of-thought · CoT · CoT prompting

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.