reasoning model

basicwidespread

A language model that reasons step-by-step before answering, making it slower but more reliable for complex questions.

In plain terms

The traditional chat model responds like someone who blurts out the answer immediately from memory. The reasoning model is more like a student who, during a test, first writes a draft: it breaks down the task, tries a solution, checks it, and only then writes the final version.

This "thinking time" can last from a few seconds to several minutes. In return, the model performs better where multiple steps must be linked: in mathematical and logical tasks, programming, analyzing long documents, and planning.

Simple questions (such as drafting an email or translating a word) do not need longer thinking: there it only makes the answer slower and more expensive. Most services today therefore decide themselves when the model should think longer, or leave the choice to you with a switch.

How it works

It receives your question and does not start writing the final answer but creates an internal "draft": it breaks the problem down into subtasks.
In this draft, it proceeds step by step: formulating assumptions, calculating, comparing options, and if it encounters a contradiction, it steps back and tries something else.
At the end of the reasoning process, it checks the result, then writes the shorter, organized answer intended for you.
You usually do not see the draft in full, only a summary; however, the computation spent counts toward response time and cost.
The capability is primarily developed through reinforcement learning: the model solves many tasks with verifiable outcomes (such as math problems or executable code) and receives a reward for correct solutions. In this way, it learns which line of reasoning leads to the goal.

An everyday example

You ask the chatbot to compare three mortgage offers over the full term with different interest rate periods. The screen displays the 'Thinking...' message, possibly with a few short summary lines about what it is working on, and after half a minute you receive a step-by-step calculation. This is when a reasoning model is at work.

Who is affected

employeeFor analysis, planning, and spreadsheet calculations, this mode is worth choosing; for routine text, waiting for it is unnecessary. → Work & HR

developerCoding agents build on reasoning models because debugging and modifications affecting multiple files require planning in multiple steps. → Code

parent, student, teacherThe derived solution is useful for learning but also solves homework in minutes, so knowledge must be assessed differently. → Education & Family

entrepreneur, executiveThinking time costs money, so it is worth deciding which tasks justify using the more expensive, slower model.

A common misconception

MythThe reasoning model thinks like a human, and what it deduces is certainly true.

In realityIt generates text just like other language models, but it first writes down intermediate steps for itself. This makes it less prone to errors, but even a convincingly reasoned answer can be wrong.

Limits and risks

It is slower and more expensive than a traditional model, and does not provide a better answer to simple questions.
It can still make mistakes and fabricate facts; moreover, due to the long derivation, the error appears more credible.
The shown reasoning is often just a summary and does not necessarily reflect exactly how the model arrived at the result.
What it does not know (breaking news, your company data) cannot be replaced by reasoning; search or attached documents are needed.

Similar, but not the same

chain-of-thought – Chain of thought was originally a questioning technique where you request step-by-step reasoning, whereas a reasoning model trained on this performs it automatically. The term is now also used for the internal reasoning process of a reasoning model.
AI-agent – A reasoning model only thinks things through and responds; an agent also uses tools and acts on its own; yet the "brain" of most agents is a reasoning model.
deep research – Deep research is a ready-made service that performs web searches and source collection, while the reasoning model is the engine working within it.

Related concepts

Broader

large language model, generative AI, foundation model

Narrower

extended thinking

On the edge

chain of thought, thinking framework, AI-agent, token, hallucination, prompt

A brief history

The benefit of step-by-step reasoning was demonstrated by Google researchers in 2022 in their study on chain-of-thought. The first widely known model specifically trained for this was OpenAI's o1, presented in September 2024. From the end of 2024 and throughout 2025, major developers released their own versions, including the open DeepSeek-R1.

Where things stand

By the end of 2026, almost all leading models can work in a reasoning mode, and most chat applications switch between fast and thoughtful answers automatically or with a toggle. It has become the default engine of coding and office agents.

Status: 2026 November

Also known as

reasoning model · reasoning model · reasoning model · reasoning model · reasoning language model · thinking model · extended thinking · RLM

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.