---
title: "Reasoning model: definition and example – Carl-Uwe Würfel"
description: "Reasoning models explained: how a language model works through a chain of thought before it answers, what thinking tokens cost, and when they pay off."
url: https://carl-wuerfel.de/en/glossary/reasoning-modell
author: Carl-Uwe Würfel
---

1. [Home](https://carl-wuerfel.de/en)
  2. [Knowledge](https://carl-wuerfel.de/en/knowledge)
  3. [Glossary](https://carl-wuerfel.de/en/glossary)
  4. Reasoning model

# Reasoning model

A reasoning model is a language model that generates a longer internal chain of thought before it answers, spending more compute on every request than a standard model does.

## On this page

  1. How a reasoning model arrives at its answer
  2. Thinking tokens cost money even when nobody reads them
  3. You can set how long the model thinks
  4. A reasoning model pays off on tasks with many dependent steps
  5. The chain of thought is not a reliable explanation
  6. Where the term stops working
  7. Frequently asked questions
  8. Sources

The whole chain of thought is billed, including the part nobody sees.

_Figure: A request to a reasoning model in three steps. The request comes in, the model produces thinking tokens in a thinking phase, breaking the task down, checking solution paths and discarding wrong ones, and then writes the answer. At the end comes the bill: thinking tokens and the answer both count as output tokens. Below, a note that users usually see only the answer or a summary, while the full thinking phase is billed._

A reasoning model answers in two phases. First it produces intermediate steps: it breaks the task down, tries a solution path, checks it and switches approach if the path doesn’t hold. Then it writes the answer. Those steps are [tokens](https://carl-wuerfel.de/en/glossary/token) like any other output, though users usually see none of them or only a summary. Technically it’s still a [large language model](https://carl-wuerfel.de/en/glossary/llm) (LLM), one that training has taught to produce this [chain of thought](https://carl-wuerfel.de/en/glossary/chain-of-thought) by itself.

In September 2024 OpenAI introduced o1, a model that, in the company’s words, thinks before it answers. Its performance rose with more [reinforcement learning](https://carl-wuerfel.de/en/glossary/reinforcement-learning) in training and with more thinking time per request. Compute spent at answer time has come to be called test-time compute. In September 2025 the DeepSeek team reported in Nature that reinforcement learning can bring out these abilities without humans writing out solution paths.

For executives it comes down to cost and wait time. Depending on the task, OpenAI says a model may generate from a few hundred to tens of thousands of thinking tokens, and OpenAI, Anthropic and Google bill them as output. Whether that’s worth it depends on the task, and you can measure it.

Example

A machinery manufacturer has AI pre-screen incoming supplier contracts. A fast model without a thinking phase sorts every contract and flags the standard ones. Only contracts with unusual liability or payment clauses go to a reasoning model at a high effort level, which compares the clauses with internal policies and explains its assessment. The legal team checks that assessment against the contract text and decides; thinking tokens and turnaround time are tracked per contract.

Chapter 1 of 6

## How a reasoning model arrives at its answer.

A standard language model writes its answer directly, token by token. A reasoning model first runs a thinking phase and reads its own steps back as it writes.

It learns this mainly through reinforcement learning. The model gets tasks with checkable results, such as math or code with tests, and is rewarded when the final answer is right. OpenAI says o1 learns this way to spot its own mistakes, break hard steps down and try another approach when needed. DeepSeek reports that R1 developed self-checking and strategy changes on its own during training.

Reasoning models are therefore strongest where a right answer can be verified: math, programming, science questions, processes with many dependent steps. On tasks without a clear right or wrong, the edge is often smaller. OpenAI itself found in 2024 that human raters didn’t prefer o1-preview on some natural-language tasks.

Chapter 2 of 6

## Thinking tokens cost money even when nobody reads them.

OpenAI’s developer documentation says reasoning tokens stay invisible through the API but take up space in the [context window](https://carl-wuerfel.de/en/glossary/context-window) and are billed as output tokens. At Anthropic you pay for the full thinking phase, not the summary that appears in the response. Google bills Gemini’s output and thinking tokens together.

The whole chain of thought is billed, including the part nobody sees. The length of visible answers is no basis for cost planning.

Anthropic reports separately in each response’s usage data how many output tokens went into thinking. Track that by task type, and after a test phase you know what a contract review actually costs. A cap on output per request catches outliers; at Anthropic it covers thinking and answer combined, and if it’s too tight the answer gets cut off.

Chapter 3 of 6

## You can set how long the model thinks.

At OpenAI the dial is called reasoning effort: low settings save time and tokens, high ones make the model think more thoroughly. Anthropic started with a fixed token budget for the thinking phase (extended thinking). On newer Claude models the model decides whether and how long to think, steered by an effort level; at a low level it skips thinking on simple questions. Google controls Gemini with a thinking level.

The level is a setting per task, not a fixed property of the model. Anthropic notes that less thinking can lower quality and recommends measuring on a representative sample before production how usage, latency and quality change. A higher level lengthens the solution path; it doesn’t guarantee a correct result.

Chapter 4 of 6

## A reasoning model pays off on tasks with many dependent steps.

OpenAI lists complex problem solving, coding, scientific reasoning and multi-step agent workflows as typical uses. Google recommends little thinking for simple queries and a lot for demanding coding, math and planning tasks. By both accounts, the extra cost pays off when steps depend on each other and a mistake costs more than the wait.

In a company, that means checking a contract against internal policies or tracking down an error in an analysis. A website chat is a poorer fit because speed counts there. In high-volume work, every thinking token gets multiplied.

One possible setup combines both: a fast model without a thinking phase pre-sorts, and only hard cases go to a reasoning model. That routing needs tested rules. If they’re too narrow, hard cases end up with the fast model.

Chapter 5 of 6

## The chain of thought is not a reliable explanation.

Many interfaces show a summary of the thinking phase. It reads like a justification, but it isn’t a reliable one. In 2025 Anthropic tested whether reasoning models disclose that they used a hint slipped to them. Claude 3.7 Sonnet mentioned the hint 25 percent of the time on average, DeepSeek R1 39 percent of the time. The researchers concluded that you can’t always rely on what models say about how they reasoned. OpenAI doesn’t show the raw chain of thought at all; users see a summary the model writes.

The chain helps with debugging, but it doesn’t work as evidence for customers, auditors or regulators. The result gets checked against the source, such as the contract text or the invoice. A reasoning model can still make up facts, because thinking longer doesn’t supply knowledge it lacks.

Chapter 6 of 6

## Where the term stops working

A reasoning model computes longer, but it knows no more than its training and context provide; current or internal knowledge comes in through [RAG](https://carl-wuerfel.de/en/glossary/rag) or tools. Chain-of-thought prompting is an instruction to any model, whereas a reasoning model learned step-by-step thinking in training. An AI agent can use a reasoning model for planning but also acts in other systems through tools.

## Frequently asked questions

What is a reasoning model?

A reasoning model is a language model that generates an internal chain of thought before it answers. It breaks the task down, checks solution paths and only then writes the result. That mainly helps on multi-step tasks and costs more time and tokens.

What is the difference between a reasoning model and a regular language model?

A regular language model writes its answer directly, while a reasoning model runs a thinking phase first. Many current models can run with or without a thinking phase.

What is the difference between a reasoning model and chain-of-thought prompting?

Chain-of-thought prompting is an instruction in the prompt, while a reasoning model learned to work step by step in training. With prompting, you ask any model to write out its steps, as Jason Wei and colleagues described in 2022. A reasoning model does it by itself.

Why are reasoning models more expensive and slower?

They generate extra thinking tokens before the answer, which take compute time and are billed as output. According to OpenAI, a task can produce from a few hundred to tens of thousands of these tokens, even when the visible answer is short.

When should you use a reasoning model?

A reasoning model pays off on multi-step tasks where a mistake costs more than the wait, such as contract review or debugging. For customer chats and large volumes of routine cases, a fast model is usually better. A test with your own cases settles it.

Which reasoning models are there?

Well-known examples are OpenAI o1 from 2024 and DeepSeek-R1, whose weights are freely available under the MIT license. Anthropic and Google offer an adjustable thinking phase in Claude and Gemini.

Can you trust the chain of thought a reasoning model displays?

Only to a limited extent. In a 2025 Anthropic study, reasoning models disclosed a hint they had used in only some cases. The chain helps with debugging; the result gets checked against the source.

## Further reading

  * [**AI glossary**](https://carl-wuerfel.de/en/glossary)
  * [**AI automation**](https://carl-wuerfel.de/en/ai-automation)
  * [**AI cost management: know the cost per call or fly blind**](https://carl-wuerfel.de/en/notes/controlling-ai-api-costs)
  * [**Chain of thought (CoT)**](https://carl-wuerfel.de/en/glossary/chain-of-thought)
  * [**AI inference**](https://carl-wuerfel.de/en/glossary/inferenz)
  * [ Term**LLM: Large Language Model**](https://carl-wuerfel.de/en/glossary/llm)
  * [ Term**Token**](https://carl-wuerfel.de/en/glossary/token)
  * [ Term**AI benchmark**](https://carl-wuerfel.de/en/glossary/ki-benchmark)

## Sources

  1. [OpenAI: Learning to reason with LLMs, September 12, 2024](https://openai.com/index/learning-to-reason-with-llms/)
  2. [OpenAI: Reasoning models, developer documentation, as of November 2026](https://developers.openai.com/api/docs/guides/reasoning)
  3. [Anthropic: Steering thinking, Claude Platform documentation, as of November 2026](https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost)
  4. [Google: Gemini thinking, Gemini API documentation, as of November 2026](https://ai.google.dev/gemini-api/docs/thinking)
  5. [DeepSeek-AI (Daya Guo et al.): DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning, Nature 645, September 2025](https://doi.org/10.1038/s41586-025-09422-z)
  6. [DeepSeek-AI: DeepSeek-R1, model card and license, Hugging Face, as of November 2026](https://huggingface.co/deepseek-ai/DeepSeek-R1)
  7. [Jason Wei et al. (Google Research): Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, NeurIPS 2022](https://arxiv.org/abs/2201.11903)
  8. [Anthropic: Reasoning models don’t always say what they think, April 3, 2025](https://www.anthropic.com/research/reasoning-models-dont-say-think)
