Is chain-of-thought the same thing as a "reasoning model"?

Not quite. Chain-of-thought is the technique — writing out reasoning steps before an answer. A reasoning model is an AI trained to generate that reasoning automatically, without being prompted to.

How do I get an AI to use chain-of-thought reasoning?

Ask it directly — adding a phrase like “think step by step” or asking it to show its work often improves accuracy on hard questions. Many current models also have a built-in extended-reasoning or “thinking” mode you can switch on.

Does showing its reasoning make an AI answer more trustworthy?

It helps, but it isn’t proof. Research has shown a model’s written-out steps don’t always match what it actually computed internally, so a chain of thought is worth checking, not blindly trusting.

Who invented chain-of-thought prompting?

It was introduced by Google researchers in a 2022 paper, “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.”

Does chain-of-thought reasoning cost more to use?

Yes — every reasoning step the model writes is made of tokens, and those are generated and billed like any other output, so longer reasoning means higher cost and slower answers.

What Is Chain-of-Thought Reasoning — and How Does AI 'Think'?

Chain-of-thought reasoning is a technique that gets an AI model to work through a problem in visible, step-by-step language — closer to how a person reasons out loud — instead of jumping straight to a final answer. First described by researchers in 2022, it turned out to be one of the simplest, most effective ways to make large language models better at math, logic, and multi-step problems, and it now sits at the core of how today’s advanced “reasoning” AI models work.

What chain-of-thought actually means

A standard large language model generates its answer one token at a time, and each new token is chosen based on everything written so far — including the model’s own prior tokens. Chain-of-thought prompting, one of the foundational techniques in prompt engineering, exploits that fact: if the model is encouraged to write out its intermediate reasoning steps first — “first I need to work out X, then Y, so the answer is Z” — those steps become part of the context it conditions on, and the final answer tends to be far more accurate.

The original 2022 paper, “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” showed that adding a handful of worked examples with reasoning steps to a prompt let large models solve grade-school math problems and multi-step logic puzzles they had previously gotten wrong. A simpler variant, zero-shot chain-of-thought, needs no examples at all — appending the phrase “let’s think step by step” to a question was enough to noticeably raise accuracy on many benchmarks. The effect mostly shows up in large models; smaller ones often don’t benefit, and can even get worse.

From a prompting trick to a built-in skill

Chain-of-thought started as something a user typed into a prompt. It has since become something reasoning models do automatically, without being asked. OpenAI’s o1, released in 2024, was the first widely used model trained specifically to generate an extended internal chain of reasoning — often called “reasoning tokens” — before producing a visible answer. Rather than only rewarding a correct final answer, its training also rewards correct intermediate steps, using reinforcement learning to make the reasoning itself more reliable. Google’s Gemini and Anthropic’s Claude have since shipped their own extended-reasoning modes built on the same underlying idea.

This is closely related to test-time compute: letting a model spend more computation — more reasoning steps — on a hard question before answering, rather than simply making the model bigger. A longer chain of thought effectively trades speed and cost for accuracy, which is also why reasoning-model answers tend to cost more and take longer to arrive: every reasoning token gets generated, and billed, like any other output token.

Why it matters — and where it breaks down

The practical payoff is real: models that reason step by step do measurably better on math competitions, coding problems, and multi-step logic than the same model asked to answer directly. Being able to see the steps can also make an answer easier for a person to check — if a chain of reasoning contains an obvious error, a reader can catch it before trusting the conclusion.

But the visible reasoning is not always a faithful record of what the model actually did internally. Interpretability research has found cases where the written-out steps look plausible but don’t fully match the computation behind the final answer — the model can, in effect, produce a reasonable-sounding explanation after the fact rather than a genuine trace of its process. That makes chain-of-thought a highly useful tool for improving accuracy and for giving a rough window into a model’s reasoning, but not, by itself, a guarantee that an answer is correct or that the explanation is honest.

For anyone using an AI assistant, the practical takeaway is simple: for a hard, multi-step question, picking a model built for extended reasoning or explicitly asking it to work step by step — you can try the technique yourself with nothing more than a prompt — is one of the most reliable ways to get a better answer. Just treat the reasoning it shows as a useful check, not as proof.

What chain-of-thought actually means

From a prompting trick to a built-in skill

Why it matters — and where it breaks down

Frequently asked questions