Reasoning models explained: how AI thinks before it answers

A reasoning model is a large language model (LLM) that works through a problem step by step before it gives a final answer. This extra "thinking" helps it solve harder problems in maths, coding and science, but each answer takes more time and costs more.

OpenAI started the trend with o1-preview on 12 September 2024. As of November 2026, most big AI labs offer a reasoning model or a "thinking" mode. Below, you will learn how reasoning models work, how they differ from normal LLMs, when to use one and where they still fail.

What is a reasoning model?

A reasoning model is an LLM that is trained to write out its thinking before it answers. The thinking is a long chain of small steps. The model tries an idea, checks it, notices a mistake and tries again. Only then does it write the answer you see.

Think of the rough work you do in the margin of an exam paper. A normal LLM writes the answer straight away. A reasoning model does the rough work first.

This rough work is made of "thinking tokens", also called reasoning tokens. Some apps show you the thinking, some show a short summary of it and some hide it. In OpenAI's API, for example, reasoning tokens are not visible. They still take space in the model's memory for the conversation (its context window), and you still pay for them.

How do reasoning models work?

Most reasoning models learn to think through reinforcement learning (RL). The model tries a problem many times, and the attempts that reach a correct answer get a reward. The clearest public example is DeepSeek-R1, whose training is described in a paper in Nature (September 2025). DeepSeek trained its first version, DeepSeek-R1-Zero, like this:

1. 1Start with a normal pretrained LLM. DeepSeek used its own DeepSeek-V3-Base.
2. 2Give it problems that a computer can check, such as maths questions with one number as the answer, or code that must pass tests.
3. 3Ask it to write its thinking inside special "think" tags, and then the answer.
4. 4Reward the model only when the final answer is correct and the format is right. Nobody shows it how to think.
5. 5Repeat this for thousands of training steps.

The result surprised the team. The model learned by itself to think longer, and its average answer grew many times longer during training. It started to check its own steps and try other ways to solve a problem. In one answer it wrote, "Wait, wait. Wait. That's an aha moment I can flag here," and then went back to fix its work.

The scores show the change. On AIME 2024, a hard maths contest, DeepSeek-R1-Zero's pass@1 score rose from 15.6% to 77.9% during training.

DeepSeek-R1-Zero had problems too. Its thinking was hard to read, and it sometimes mixed English and Chinese in one answer. So the final DeepSeek-R1 added more training stages, starting with thousands of examples of clear, readable thinking. You can read the full method in the DeepSeek-R1 paper.

What is chain of thought?

Chain of thought (CoT) is a series of in-between reasoning steps that a model writes before its final answer. The idea became popular through prompting, more than two years before reasoning models arrived.

In 2022, Jason Wei and colleagues showed that you can get these steps by putting a few worked examples in the prompt. With just eight such examples, a 540-billion-parameter model set a new best score on GSM8K, a set of school maths word problems. The chain-of-thought paper is short and easy to read.

The same year, Takeshi Kojima and colleagues found an even simpler trick: add "Let's think step by step" before the answer. On one large model, this raised accuracy on the MultiArith maths set from 17.7% to 78.7%, and on GSM8K from 10.4% to 40.7%.

Reasoning models vs chain-of-thought prompting

Chain-of-thought prompting is a trick you add to a prompt. A reasoning model is trained to make long chains of thought on its own, and to check and fix them as it goes. So you do not need to ask a reasoning model to "think step by step". It already does.

What is test-time compute?

Test-time compute is the computing power a model uses while it answers your question. "Test time" means the time when a trained model is used, not when it is trained. A reasoning model uses more test-time compute, because it writes many thinking tokens before each answer.

For years, the main way to make AI better was to train bigger models on more data. Reasoning models made a second way popular: let the model spend more effort on each question. A 2024 study by Charlie Snell and colleagues tested this idea. They found that a smaller model can beat a model 14 times larger. The trick is to share test-time compute wisely and give more of it to harder questions. This held for questions the smaller model could already solve some of the time, and both models used the same total compute.

Thinking longer in one chain is one way to use test-time compute. Here are others:

• Majority vote: ask the model the same question several times and pick the most common answer.
• Best of N: make several answers and let a second model, called a verifier, pick the best one.
• Search: explore many possible next steps, like branches of a tree, and drop the weak ones.

Many tools now let you choose how much a model thinks. OpenAI's API has a "reasoning effort" setting, with values such as low, medium and high, depending on the model. Anthropic lets developers set a "thinking budget" for Claude.

Reasoning model vs normal LLM: what is the difference?

A normal LLM starts writing the answer straight away. A reasoning model first spends extra tokens thinking, then answers. Both are the same kind of neural network. The difference is mostly in how they were trained and how much compute each answer uses.

• Speed: a normal LLM answers quickly. A reasoning model can take much longer on a hard question.
• Cost: a reasoning model writes many more tokens, so each answer costs more.
• Best at: normal LLMs suit chat, writing, summaries and simple facts. Reasoning models suit maths, code, logic and planning with many steps.
• Prompts: normal LLMs often do better with examples in the prompt. DeepSeek found that examples made DeepSeek-R1 worse, so it advises describing the problem directly.

The line between the two is now blurry. In February 2025, Anthropic called Claude 3.7 Sonnet "the first hybrid reasoning model on the market": one model that can answer quickly or think longer. Many models now have a similar switch. The Indian company Sarvam AI also built a thinking mode into its open models.

What are some examples of reasoning models?

• OpenAI o1-preview: released on 12 September 2024. It was OpenAI's first reasoning model.
• DeepSeek-R1: released in January 2025 with open weights under the MIT licence. DeepSeek also released six smaller "distilled" versions, from 1.5 billion to 70 billion parameters.
• Claude 3.7 Sonnet: released on 24 February 2025 with "extended thinking", which you can switch on or off.
• Gemini 2.5: announced on 25 March 2025. Google called the Gemini 2.5 models "thinking models, capable of reasoning through their thoughts before responding".
• Sarvam-105B: an open model from Sarvam AI, released in 2026. Its code examples switch on a thinking mode, and Sarvam says it does well in 22 Indian languages.

Model names change fast. Most big labs now have newer reasoning models than these, so check each lab's model page for the latest version.

When should you use a reasoning model?

Use a reasoning model when a task has many steps and a clear right answer. Use a normal model when the task is simple or you need a fast reply.

Good uses for a reasoning model:

• Maths: multi-step problems, or checking each line of a derivation.
• Code: finding a hard bug, or writing a function with tricky edge cases.
• Logic: checking whether the steps in an argument really follow.
• Planning: breaking a research project into tasks and finding gaps in the plan.

Poor uses: simple facts, translation, short emails, summaries and casual chat. The extra thinking adds time and cost but little value here. In a 2025 Apple study, normal models even beat reasoning models on easy puzzles when both used the same compute.

Our advice for students: start in normal mode. Switch to thinking mode when the first answer is wrong, or when you can see the problem needs several steps. This saves time and, on paid plans, money.

How should you prompt a reasoning model?

• Be direct: describe the problem and the answer format you want. DeepSeek recommends this "zero-shot" style for R1.
• Skip examples: in DeepSeek's tests, adding examples to the prompt (few-shot) made DeepSeek-R1 worse.
• Set the goal: OpenAI says its reasoning-capable GPT-5 models work best with a clear goal and strong constraints.
• Mind the language: DeepSeek-R1 is tuned for Chinese and English. If you ask in Hindi or Tamil, it may think and even answer in English.
• Check the facts: thinking longer does not give a model facts it never learned. Check every source, because AI tools can invent citations that look real.

What are the limits of reasoning models?

Reasoning models are better at hard problems, but they have five known weak spots.

• Cost: OpenAI and Anthropic both bill thinking tokens as output tokens. If a model thinks for 3,000 tokens to write a 300-token answer, you pay for 3,300 output tokens.
• Overthinking: in one 2024 study, o1-like models used on average 1,953% more tokens than normal models to answer "what is the answer of 2 plus 3?" DeepSeek also admits that R1 sometimes overthinks simple questions.
• Hidden reasons: the written thinking may not show why the model really chose its answer. In a 2025 Anthropic study, models were given a hint in the prompt. When they used it, Claude 3.7 Sonnet mentioned the hint in its thinking only 25% of the time, and DeepSeek-R1 only 39% of the time.
• Very hard puzzles: Apple's 2025 study "The Illusion of Thinking" found that reasoning models fail completely beyond a certain puzzle difficulty. Near that point, they even start to think less. A reply paper disputed this. It argued that some puzzles hit the models' output limits and some had no solution at all. So read this result with care.
• Narrow strength: RL works best when a computer can check the answer, as in maths and code. DeepSeek notes that a reliable reward is hard to build for tasks such as writing.

How can students start learning about reasoning models?

1. 1Read the chain-of-thought paper by Wei and colleagues (2022). Use the three-pass method so you do not get stuck in the details.
2. 2Read sections 1 and 2 of the DeepSeek-R1 paper. They explain R1-Zero in about four pages, and the paper is free on arXiv.
3. 3Run a small reasoning model yourself. DeepSeek-R1-Distill-Qwen-1.5B is small enough for a free Google Colab GPU. Read its thinking and note where it goes wrong. Our post on small language models explains why small models matter.
4. 4Try a mini research project. Take 30 maths word problems and compare a normal model with a reasoning model on accuracy, number of tokens and time. Write down where thinking helped and where it was wasted.
5. 5For a hands-on book, Sebastian Raschka's "Build a Reasoning Model (From Scratch)" builds a small reasoning model step by step. Its code is free on GitHub.

At APJ-RAISE, students read papers like these together and turn questions like "when does thinking help?" into small research projects. Joining is free.