A reasoning model is a kind of large language model (LLM): one trained to write hidden working before it answers. An ordinary LLM starts its reply straight away. So the difference is one extra step, not a different technology. Reasoning models think first, which makes them slower and makes them use more tokens, but stronger on hard, many-step problems such as maths, code and planning.
- An LLM writes text by predicting what comes next, piece by piece. It is the engine inside every AI chatbot.
- A reasoning model is an LLM that writes hidden “reasoning tokens” before the answer you see.
- Use reasoning for puzzles, code, maths and plans with many parts. Skip it for quick questions, where it only adds waiting.
- Prompt it differently: state the goal and limits clearly; you rarely need “think step by step”.
- The visible “thinking” is not a reliable window into how it got the answer. Check the answer itself.
| Ordinary LLM | Reasoning model | |
|---|---|---|
| What it is | A language model that predicts the next piece of text. | The same kind of model, trained to work through a problem first. |
| How it answers | Starts the reply at once. | Writes hidden working, then the reply. |
| Speed and tokens | Faster, fewer tokens. | Slower, more tokens, set by an effort level on some models. |
| Good for | Drafting, rewriting, summarising, quick questions. | Complex problems, coding, science, multi-step tasks. |
| Prompting | Asking for steps can help. | Say what you want; the steps happen anyway. |
Think of two students taking a maths exam. One writes the answer straight down. The other fills a sheet of scrap paper first, trying ideas and crossing some out, then writes the answer. Same student, same knowledge; the second just took time to work it through. You mostly see only the answer, not the scrap paper.
What is an LLM?
The plain point: a model that has learned the patterns of language and writes by predicting the next piece of text.
Google’s training course describes a language model as one that “estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens.”1 Tokens are the small pieces of text a model reads and writes. Anthropic puts it in five words: “Claude writes text one word at a time.”2
That is the engine inside ChatGPT, Claude and Gemini. More on LLM meaning.
What makes a reasoning model different?
The plain point: it writes working for itself before it writes for you, and it was trained to make that working useful.
OpenAI, which makes several of these models, says: “Reasoning models use internal reasoning tokens before producing a response.” Those are extra: “Reasoning models introduce reasoning tokens in addition to input and output tokens.”3
The skill comes from extra training. OpenAI said of its first reasoning model that “Our large-scale reinforcement learning algorithm teaches the model how to think productively using its chain of thought”.4 Reinforcement learning means learning by reward: good results are rewarded, so the model does more of what produced them. The AI lab DeepSeek reported that “the reasoning abilities of LLMs can be incentivized through pure reinforcement learning”.5
Note the words both labs use: reasoning “with LLMs” and reasoning abilities “of LLMs”. A reasoning model is not the opposite of an LLM. It is an LLM with a trained habit of working things through. More on reasoning model meaning.
Is it always a separate model?
The plain point: less and less. Many models now let you turn the thinking up, down or off.
OpenAI’s reasoning effort setting can include values from “none” up to “max”, depending on the model, and the trade is spelled out: “Lower effort favors speed and lower token usage.” It suggests “none” for “Latency-critical tasks that do not benefit from any reasoning.”3 Anthropic’s guide for Claude also talks about thinking being on or off: “When thinking is off, you can still encourage step-by-step reasoning by asking Claude to think through the problem.”6
So in practice “LLM or reasoning model?” is often a setting in one app, not a choice between two products. Latency, in the quote above, is the wait before an answer.
When is each one the better choice?
The plain point: reasoning for hard problems; ordinary answers for everything quick.
OpenAI says reasoning models “work especially well for complex problem solving, coding, scientific reasoning, and multi-step agentic workflows.”7 The hidden working also has a cost: OpenAI says reasoning tokens “count toward output usage and are billed as output tokens.”8
- Reach for reasoning: a tricky plan with competing limits, a bug you cannot find, a sum with several steps.
- Skip it: a word that means “happy”, a quick rewrite, a short summary. The pause buys you nothing.
That split is this site’s reading of the vendors’ guidance.
Does it change how you write prompts?
The plain point: yes. Tell a reasoning model what you want, not how to think.
OpenAI says: “Since these models perform reasoning internally, prompting them to ‘think step by step’ or ‘explain your reasoning’ is unnecessary.”9 With an ordinary model, or with thinking switched off, asking for steps can still help, as the Anthropic line above says. Prompting myths checks the other old tricks.
Can you trust the thinking it shows?
The plain point: not as proof. The working may not be how the answer was really reached.
In a 2025 test, researchers slipped models a hint and checked whether the written reasoning admitted using it. Anthropic reported that “Claude 3.7 Sonnet mentioned the hint 25% of the time, and DeepSeek R1 mentioned it 39% of the time.”10 OpenAI, for its part, says “we don’t expose the raw reasoning tokens emitted by the model”, so what an app shows may be a summary.3
A reasoning model is an LLM that thinks before it answers. Use it when the problem has many steps, state the goal plainly, and judge the answer, not how convincing the working looks.
Questions people ask
Is a reasoning model an LLM?
Yes. OpenAI and DeepSeek both describe reasoning as something taught to LLMs through reinforcement learning (sources above). It is a large language model with an extra trained step.
Why is a reasoning model slower?
Because it writes hidden reasoning tokens before the answer, on top of the answer itself (source above). OpenAI’s lower effort settings favour speed and fewer tokens.
Are reasoning models always right?
No. They are stronger on hard problems, but they can still be confidently wrong, and their shown working can leave out what really drove the answer (source above). See hallucination.
Is reasoning the same as chain of thought?
Close. Chain of thought is the step-by-step working itself; OpenAI says reinforcement learning teaches its reasoning models to use their chain of thought (source above).
Where to go next
- “Give me each word on its own.” Read LLM, reasoning model and chain of thought.
- “Which model should I use?” See choosing an AI model.
- “How do I ask each one well?” Read prompting for LLMs.
- “What else do people get wrong?” See AI terms people get wrong.
- “What happens when it runs?” See inference.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. How does the page describe a reasoning model?
- A completely different technology from an LLM
- An LLM trained to write hidden working before it answers
- A search engine that looks up answers instead of writing them
Show the answer
Answer: An LLM trained to write hidden working before it answers. The page says a reasoning model is a kind of LLM with one extra step: it thinks before it answers.
2. For which job does the page suggest you can skip reasoning?
- A quick rewrite or a word that means happy
- Finding a bug you cannot spot
- A plan with several competing limits
Show the answer
Answer: A quick rewrite or a word that means happy. The page says reasoning earns its time on hard, many-step problems and only adds waiting on quick questions.
3. What does the page say about the thinking a reasoning model shows?
- It is an exact record of how the answer was reached
- It proves the answer is right
- It may not reflect how the answer was really reached, so check the answer
Show the answer
Answer: It may not reflect how the answer was really reached, so check the answer. The page cites a test where models often did not mention a hint they had used, so the working is not proof.
- 1Google, Introduction to Large Language Models, updated 9 Jan 2026, read at source 1 Oct 2026 (as cited on this site’s LLM meaning page): “A language model estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens.”↑
- 2Anthropic, Tracing the thoughts of a large language model, 27 Mar 2025, read at source 1 Oct 2026 (as cited on this site’s LLM meaning page): “Claude writes text one word at a time.”↑
- 3OpenAI, Reasoning models, undated guide, read at source 19 Oct 2026“Reasoning models use internal reasoning tokens before producing a response.” “Reasoning models introduce reasoning tokens in addition to input and output tokens.” ↑a“Supported values are model-dependent and can include none, minimal, low, medium, high, xhigh, and max.” “Lower effort favors speed and lower token usage”; none suits “Latency-critical tasks that do not benefit from any reasoning.” ↑b“While we don’t expose the raw reasoning tokens emitted by the model”. ↑c
- 4OpenAI, Learning to reason with LLMs, 12 Sep 2024, read at source 5 Oct 2026 (as cited on this site’s reinforcement learning meaning page): “Our large-scale reinforcement learning algorithm teaches the model how to think productively using its chain of thought in a highly data-efficient training process.”↑
- 5DeepSeek-AI et al., DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, arXiv:2501.12948, submitted 22 Jan 2025, revised 4 Jan 2026, read at source 5 Oct 2026 (as cited on this site’s reinforcement learning meaning page): “Here we show that the reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories.”↑
- 6Anthropic, Prompting best practices, undated documentation, read at source 1 Oct 2026 (as cited on this site’s chain of thought meaning page): “When thinking is off, you can still encourage step-by-step reasoning by asking Claude to think through the problem.”↑
- 7OpenAI, Reasoning models, read at source 1 Oct 2026 (as cited on this site’s reasoning model meaning page): they “work especially well for complex problem solving, coding, scientific reasoning, and multi-step agentic workflows.”↑
- 8OpenAI, Understanding and counting tokens, read at source 13 Oct 2026 (as cited on this site’s AI terms people get wrong page): “Reasoning tokens are not visible as answer text, but they count toward output usage and are billed as output tokens.”↑
- 9OpenAI, Reasoning best practices, undated guide, read at source 1 Oct 2026 (as cited on this site’s chain of thought meaning page): “Since these models perform reasoning internally, prompting them to ‘think step by step’ or ‘explain your reasoning’ is unnecessary.”↑
- 10Anthropic, Reasoning models don’t always say what they think, 3 Apr 2025, read at source 13 Oct 2026 (as cited on this site’s AI terms people get wrong page): “Claude 3.7 Sonnet mentioned the hint 25% of the time, and DeepSeek R1 mentioned it 39% of the time.”↑
Reused from this site’s meaning pages and articles, with the read dates recorded there: Google’s LLM course and Anthropic’s tracing post (LLM meaning, 1 Oct 2026); OpenAI’s Learning to reason with LLMs and the DeepSeek-R1 paper (reinforcement learning meaning, 5 Oct 2026); OpenAI’s reasoning best practices and Anthropic’s prompting best practices (chain of thought meaning, 1 Oct 2026); OpenAI’s tokens help page and Anthropic’s faithfulness study (AI terms people get wrong, 13 Oct 2026).
The comparison on this page (two students and scrap paper), the quick-question examples and the advice marked “this site’s reading” are this site’s own, 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