MMOOZZGG.com

Glossary

What is a reasoning model

In short

A reasoning model is a language model that writes out its own steps of the solution before it answers, and spends more time and computing on a hard task.

Language models · reasoning model

MMOOZZGG.com editorsChecked: 1 min read

For a child

When the teacher gives a hard problem, some kids shout out the answer at once and are often wrong. Others take scrap paper, write down what they know, count step by step and check themselves. Their answer comes later, but it is right. Some AI models also write on scrap paper first and only then answer.

In plain words

A standard language model answers at once, word by word. A reasoning model first talks the solution through to itself: it splits the task into steps, tries an option, notices a mistake and goes back. This draft is sometimes longer than the answer, and the user sees all of it, a summary or nothing at all. The harder the question, the longer the model thinks. Such models are stronger in math, programming, logic and tasks where many conditions must be matched. On a simple question, like translating a word, they waste time.

For business

These models take on tasks that used to need an analyst: check a contract against a policy, find an error in a calculation, plan the multi-step work of an agent. But you also pay for the draft thinking: each hidden token is billed as part of the answer. So one request can cost several times more and take tens of seconds. For a support chat this is often too much. A sensible setup sends simple questions to a fast model and hard ones to a reasoning model. Before you choose, compare price and response time on ten of your typical tasks, not by general rankings. Prices of different models per million tokens are collected in one table.

For a specialist

A reasoning model, also called a large reasoning model, is an LLM trained to produce a long chain of thought before the final answer and to scale computing at the moment of the answer (test-time compute). The key method is reinforcement learning with a verifiable reward, on tasks where correctness can be checked automatically: math, code, tests. The OpenAI o1 system card (2024) describes a series of models trained with large-scale reinforcement learning to reason in a chain of thought before they answer. In the work of DeepSeek-AI (2025), the R1-Zero version learned to reason by pure reinforcement learning, without supervised fine-tuning on labeled examples (SFT). It developed self-checking and a change of strategy. The final DeepSeek-R1 received starter examples and multi-stage training before reinforcement learning. The skills of a large model are then moved to smaller ones by distillation. Limits: longer delay and more tokens, overthinking of simple tasks, and visible reasoning that is not a fully reliable explanation.

Example

A finance team asks a model to check whether a summary table matches the source data. A standard model says “yes” in a second. A reasoning model spends half a minute checking the rows and finds two records counted twice.

Related terms

Questions and answers

How is a reasoning model different from a standard one? A standard model answers at once. A reasoning model first builds a solution path, checks itself and only then answers.

That makes it more accurate on complex tasks, but slower and more expensive.

When should you use a reasoning model? When the task has many steps: calculations, document analysis, code, planning. For short replies to customers and simple lookups, a fast standard model is cheaper and does just as well.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us