# Reasoning model

> An AI model that works through a problem step by step before it answers.

- Level: Intermediate
- Topics: AI
- URL: https://tidbit.legios.com.ar/reasoning-model/
- In Spanish: [Modelo de razonamiento](https://tidbit.legios.com.ar/es/reasoning-model/)

A **reasoning model** is an LLM trained to think out loud before replying. Instead of producing the answer right away, it first writes a long chain of intermediate steps: it breaks the problem apart, tries an approach, checks the result, and backs up when something does not add up. Only then does it write the final answer.

The idea took off in late 2024 with OpenAI's o1, followed by DeepSeek-R1, Claude with **extended thinking** and others. These models are usually trained with reinforcement learning on problems whose answers can be checked, like math or code that has to pass tests, so the model learns which ways of thinking actually lead to correct results.

The gain is real on multi-step work: math, programming, planning, puzzles, careful analysis of a long document. The cost is also real: all that thinking is made of tokens, so answers take longer and cost more. Many APIs let you set a thinking budget, and for a quick question or a simple rewrite a regular model is often the better choice.

## An analogy

The difference between blurting out an answer in a quiz and taking a sheet of scrap paper first. The scrap paper slows you down, but on a hard problem it is what gets you to the right result.

## Keep in mind

- Writes intermediate reasoning before the final answer.
- Usually trained with reinforcement learning on checkable problems.
- Better at math, code and multi-step tasks; slower and pricier.
- Thinking tokens are billed, so many APIs let you cap them.

## Go deeper

- [Reasoning language model on Wikipedia](https://en.wikipedia.org/wiki/Reasoning_language_model)
- [Extended thinking (Anthropic docs)](https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking)
- [Reasoning models (OpenAI docs)](https://platform.openai.com/docs/guides/reasoning)

## Related terms

- [LLM](https://tidbit.legios.com.ar/llm/): A neural network trained on huge amounts of text to predict the next word.
- [OpenAI](https://tidbit.legios.com.ar/openai/): The company that made ChatGPT and turned AI into a household topic.
- [DeepSeek](https://tidbit.legios.com.ar/deepseek/): A Chinese AI lab whose open-weight models showed that top AI could be built for less.
- [Claude](https://tidbit.legios.com.ar/claude/): Anthropic's family of AI models, and the assistant app built on them.
- [Reinforcement learning](https://tidbit.legios.com.ar/reinforcement-learning/): Learning by trial and error, guided by rewards for good outcomes.
- [Token](https://tidbit.legios.com.ar/token/): The small chunk of text a language model actually reads.
- [API](https://tidbit.legios.com.ar/api/): The menu of things one program lets another program ask for.