Browse tools

How AI models work

Reasoning model

In short

A reasoning model is an AI language model trained to work through a problem step by step in a hidden "thinking" stage before giving its final answer. It is slower and costs more per answer, but tends to do better on hard maths, coding and logic tasks.

Also called
Thinking model, Large reasoning model, LRM
Topic
How AI models work
Level
Beginner
First described
OpenAI's o1 models, documented in a December 2024 system card, were trained with large-scale reinforcement learning to reason using chain of thought before answering1.

An ordinary chatbot starts writing its reply straight away. A reasoning model first produces internal reasoning tokens, then the answer2. OpenAI's documentation says these tokens are not visible through its API, but they still take up space in the context window and are billed as output tokens2.

The skill is trained in, mostly with reinforcement learning. OpenAI describes its o1 models as trained with large-scale reinforcement learning to reason using chain of thought1. DeepSeek reported in January 2025 that reasoning could be built through pure reinforcement learning, without human-written reasoning examples3. During training its R1-Zero model learned to use more thinking time, and its score on the AIME 2024 maths test rose from 15.6% to 77.9%3.

Developers can usually set how hard the model thinks. OpenAI offers a reasoning effort setting, where lower effort is faster and cheaper and higher effort is more thorough2. Anthropic's newer Claude models use adaptive thinking controlled by an effort setting, while older ones take a fixed thinking budget of at least 1,024 tokens4. Google's Gemini models adjust their thinking to the difficulty of the request by default5.

The visible reasoning is not a guaranteed record of how the answer was reached. Anthropic researchers found that reasoning models often failed to mention hints they had relied on, with the reveal rate often below 20%6.

Example

Ask a reasoning model: "A train leaves Leeds at 9:40 and the journey takes 2 hours 35 minutes, with a 20-minute delay. When does it arrive?" It writes out its working privately, then gives the answer: 12:35.

Why it matters

Reasoning models are behind the recent jumps in AI maths and coding scores. They also change the cost of AI, because you pay for the thinking as well as the answer.

Questions people ask

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

No. Chain-of-thought prompting, described in 2022, asks an ordinary model to show its steps by giving it worked examples in the prompt7. A reasoning model is trained, mostly with reinforcement learning, to produce that step-by-step thinking on its own13.

Why are reasoning models slower and more expensive?

They generate many extra tokens of thinking before the answer. OpenAI and Google both bill those thinking tokens as output, even when you only see the final answer or a summary25.

Can you see what a reasoning model is thinking?

Only partly. OpenAI's API does not show raw reasoning tokens2, Claude's API returns summarised thinking4, and Gemini returns only the final answer unless thought summaries are switched on5. Even when shown, the reasoning may leave out what actually drove the answer6.

Related terms

Chain of thought Test-time compute Reinforcement learning Large language model Inference cost Chain-of-thought monitoring

Tools that use it

Sources

  1. OpenAI (2024), OpenAI o1 System Card, arXiv:2412.16720
  2. OpenAI API documentation, Reasoning models guide
  3. DeepSeek-AI et al. (2025), DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, arXiv:2501.12948
  4. Anthropic, Claude API documentation, Extended thinking
  5. Google AI for Developers, Gemini API documentation, Gemini thinking
  6. Chen et al. (2025), Reasoning Models Don't Always Say What They Think, Anthropic (paper also at arXiv:2505.05410)
  7. Wei et al. (2022), Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, arXiv:2201.11903

Back to the glossary