# What is a reasoning model? Definition and difference from a standard LLM

> A reasoning model is a large language model trained to produce internal, step-by-step thinking before giving its answer. It is more reliable on complex problems (analysis, calculation, code, agent planning), but slower and more expensive, because its thinking tokens are billed.

## Definition
The first mainstream reasoning model was OpenAI o1, released as a preview on September 12, 2024: OpenAI described it as designed to spend more time thinking before responding. On a qualifying exam for the International Mathematics Olympiad, GPT-4o solved 13% of problems, the reasoning model 83%. In January 2025, DeepSeek published R1 and showed that reinforcement learning alone can make this behaviour emerge, without human-written reasoning examples. The principle is simple. Before the visible answer, the model generates a chain of thought: it breaks down the problem, tries several approaches, checks its calculations, corrects its mistakes. These thinking tokens are largely invisible, but they occupy the context window and are billed as output tokens (OpenAI and Anthropic documentation). In 2026, the line between standard LLMs and reasoning models is blurring: flagship models decide for themselves whether to think. At Anthropic, Claude's thinking is adaptive. The model assesses each request, answers a simple question directly and thinks longer on a multi-step problem. Developers set an “effort” level (low, medium, high, xhigh, max; medium by default on Claude Opus 5.5). OpenAI offers an equivalent setting, reasoning effort, from none to max. The higher the effort, the more careful the answer, and the slower and more expensive it is. “Reasoning” remains a metaphor: the model does not think like a human and can, after long deliberation, reach a wrong answer stated with confidence.

## Example
Illustrative case: a French retail company with 600 employees uses an AI assistant for two tasks. The first is sorting 3,000 supplier emails a day (invoice, dispute, reminder). The second is preparing a monthly analysis of margin variances by product family, from several accounting exports. For email sorting, thinking effort set to the minimum gives the same quality as high effort, with answers in one to two seconds. For the margin analysis, the same model at high effort spots a unit conversion error that the no-thinking version missed. The finance department therefore keeps two settings: low effort for volume, high effort for the monthly analysis, whose extra cost stays marginal since it runs once a month.

## For a leader
“On which use cases are we paying for thinking that serves no purpose?” Asked of your CIO or your provider, this question goes straight to the bill: thinking tokens are billed as output tokens, usually the most expensive. Three points to check. First, is the effort level set per use case, or left at maximum everywhere as a precaution? Second, has the quality gap between low and high effort been measured on real cases? On a simple task (sorting, extraction, rewording), it is often zero. Third, do users accept the latency? A customer assistant that thinks for twenty seconds loses conversations. Keep long thinking for tasks where an error is costly: financial analysis, compliance checks, agent planning.

*Standard LLM vs reasoning model: what changes for the business*

|  | Standard LLM (direct answer) | Reasoning model |
| --- | --- | --- |
| How it works | Generates the answer immediately | Generates internal thinking, then the answer |
| Cost per request | Lower | Higher: thinking tokens are billed as output |
| Latency | One to a few seconds | From a few seconds to several minutes depending on effort |
| Suitable use cases | Sorting, extraction, rewording, short answers | Numerical analysis, code, compliance, agent planning |
| Main risk | Errors on multi-step problems | Extra cost and slowness when thinking is unnecessary |
| Setting | Choice of model | Effort level (from low to maximum) |

## FAQ

**What is a reasoning model in AI?**
It is a large language model trained to think step by step before answering. It generates internal thinking (breaking down the problem, trying approaches, checking), then its answer. It is more reliable on multi-step problems, slower and more expensive per request.

**What is the difference between a reasoning model and a standard LLM?**
A standard LLM produces its answer directly, token by token. A reasoning model first produces thinking tokens, billed as output tokens. In 2026, most flagship models do both: they decide for themselves whether to think, according to an effort level set by the developer.

**What are examples of reasoning models?**
OpenAI o1 (preview in September 2024) led the way, followed by DeepSeek-R1 (January 2025), released with open weights. In 2026, Anthropic's Claude Opus 5.5 includes adaptive thinking controlled by effort level, and OpenAI recommends GPT-6 Astra for most reasoning tasks.

**Do reasoning models make fewer mistakes?**
They make fewer mistakes on multi-step problems (calculation, logic, code, planning), but they are not infallible. Long thinking can end in a wrong answer stated with confidence. Source checks and human review remain necessary on high-stakes topics.

**Why does a reasoning model cost more?**
Because internal thinking consumes tokens, billed as output tokens even when you do not see them. On a hard problem, thinking can account for most of the billed tokens. The effort level is the main lever to control this cost.

**Should reasoning always be switched on?**
No. For sorting, extraction or rewording, low effort is usually enough and cuts cost and latency. Keep high effort for tasks where an error is costly, and measure the quality gap on your own cases before deciding.

## See also

- [LLM (Large Language Model)](https://wyp.agency/en/glossary/llm/)
- [Inference](https://wyp.agency/en/glossary/inference/)
- [Cost per token](https://wyp.agency/en/glossary/cost-per-token/)
- [Token](https://wyp.agency/en/glossary/token/)
- [AI agent](https://wyp.agency/en/glossary/ai-agent/)
- [Hallucination](https://wyp.agency/en/glossary/hallucination/)
- [Reinforcement learning](https://wyp.agency/en/glossary/reinforcement-learning/)
- [RLHF (reinforcement learning from human feedback)](https://wyp.agency/en/glossary/rlhf/)

## Sources

- Introducing OpenAI o1-preview, OpenAI, September 12, 2024 (https://openai.com/index/introducing-openai-o1-preview/)
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, DeepSeek-AI, arXiv, January 22, 2025 (https://arxiv.org/abs/2501.12948)
- Reasoning models (reasoning effort, billing of reasoning tokens), OpenAI documentation (https://developers.openai.com/api/docs/guides/reasoning)
- Steering thinking (effort levels, adaptive thinking, pricing), Claude documentation, Anthropic (https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost)

---
Entry from the WYP.agency AI glossary · https://wyp.agency/en/glossary/reasoning-model/ · Last updated : 2026-10-14