A reasoning model is an AI model that writes out hidden working before it gives its answer, a bit like doing rough notes first. That extra step makes it slower, but better at hard, multi-step problems.
Think of a maths exam. An ordinary model blurts out the answer straight away. A reasoning model fills a sheet of scrap paper first, trying ideas and crossing some out, then writes the final answer. You mostly see the answer, not the scrap paper.
In a sentence: “For the tax puzzle I switched to a reasoning model; it took a minute longer but caught the mistake the quick model missed.”
What the term means
“Reasoning” here means extra text the model writes for itself. AI models write in small chunks called tokens (pieces of words). A reasoning model produces a batch of hidden tokens, its working, before the reply you read. OpenAI, which makes several of these models, says they “use internal reasoning tokens before producing a response.”1
That working helps the model plan and check itself. OpenAI says the models do best on hard, many-step jobs.1
A simple example
Ask a tricky planning question and you will see the difference. Say you ask: “Plan a three-day trip for four people where one cannot walk far and two want museums.” An ordinary model starts writing a plan at once. A reasoning model may pause first, sometimes showing a “thinking” note, while it weighs the limits against each other. Then it answers.
For a quick question, such as a word that means “happy”, the pause buys you nothing. Reasoning earns its time on puzzles, code, maths and plans with many moving parts.
Why it matters to you
You prompt it differently. Old advice said to add “think step by step” to your request. A reasoning model already does that, and OpenAI says the old trick can even make results worse.2 Instead, say what you want, the limits, and what a good answer looks like. Then let it work.
The visible working is not a window into its mind. Some apps show the model’s working. Anthropic, which makes Claude, tested whether that written working is honest about how the answer was reached, and warned it may not be.3 So check the answer itself, not how convincing the working looks.
Often confused with
| Term | What it means |
|---|---|
| Thinking like a person | Not the same: the “reasoning” is generated text, not a mind. AI terms people get wrong explains the gap. |
| Chain of thought | Means: the step-by-step working itself. Reasoning models are trained to produce it on their own. |
| Always right | Not the same: reasoning models still make confident mistakes. See why AI makes things up. |
Questions people ask
What is the difference between a reasoning model and a regular AI model?
A regular model starts its reply straight away. A reasoning model first writes hidden working: OpenAI says these models “use internal reasoning tokens before producing a response”, which is why they are slower but stronger on complex problems.1
Should you tell a reasoning model to think step by step?
Usually not. OpenAI warns that telling these models to “think step by step” may not help “(and can sometimes hinder it)”. Say clearly what you want instead.2
Can you trust the “thinking” a reasoning model shows?
Not fully. Anthropic’s researchers write that “There’s no specific reason why the reported Chain-of-Thought must accurately reflect the true reasoning process”, so judge the answer, not the working.3
Learn more on this site
- Prompting myths — which old prompting tricks stopped working on reasoning models.
- Prompting for LLMs — how ChatGPT, Claude and Gemini want to be asked.
- Inference meaning — the reasoning happens while the model is running, not while it is learning.
- The AI glossary and AI for beginners — the basics, in plain words.
Related on this site: Latency meaning · Reinforcement learning meaning · LLM vs Reasoning Model
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What makes a reasoning model different?
- It writes out hidden working before it gives its answer
- It always answers instantly
- It can only do maths
Show the answer
Answer: It writes out hidden working before it gives its answer. Like filling a sheet of scrap paper before the final answer, it produces hidden working first.
2. When is a reasoning model worth the extra wait?
- For a quick question, like a word that means happy
- For hard, many-step jobs like puzzles, code, maths and plans
- Only when the internet is slow
Show the answer
Answer: For hard, many-step jobs like puzzles, code, maths and plans. For quick questions the pause buys you nothing; reasoning earns its time on problems with many moving parts.
3. Can you trust the “thinking” a reasoning model shows?
- Yes, it shows exactly how it got the answer
- Yes, as long as the working is long
- Not fully, so judge the answer itself, not the working
Show the answer
Answer: Not fully, so judge the answer itself, not the working. Anthropic warned that the written working may not honestly show how the answer was reached.
Related terms: Multimodal · Prompt chaining · Prompt engineering · Recommendation algorithm · Role prompting · Sentiment analysis
- 1OpenAI, Reasoning models, read at source 1 Oct 2026“Reasoning models use internal reasoning tokens before producing a response.” ↑athey “work especially well for complex problem solving, coding, scientific reasoning, and multi-step agentic workflows.” ↑b“Reasoning models use internal reasoning tokens before producing a response.” they “work especially well for complex problem solving, coding, scientific reasoning, and multi-step agentic workflows.” ↑c
- 2OpenAI, Reasoning best practices, read at source 22 Sep 2026 (as cited on this site’s prompting myths page): “Some prompt engineering techniques, like instructing the model to ‘think step by step,’ may not enhance performance (and can sometimes hinder it).”↑a↑b
- 3Anthropic, Reasoning models don’t always say what they think, 3 Apr 2025, read at source 1 Oct 2026“There’s no specific reason why the reported Chain-of-Thought must accurately reflect the true reasoning process”. ↑a“There’s no specific reason why the reported Chain-of-Thought must accurately reflect the true reasoning process”. ↑b
The comparisons on this page (the maths exam and scrap paper, the trip plan) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK