Reasoning model

What it means

A reasoning model is a kind of AI language model that works through a problem in steps before it gives you an answer. OpenAI's developer guide says these models "use internal reasoning tokens before producing a response." (A token is a small piece of text that the model reads or writes.) OpenAI says this helps the model "plan, use tools effectively, inspect alternatives, recover from ambiguity, and solve harder multi-step tasks."¹ Google says the same thing about its Gemini models in simpler words: "Gemini reasons internally before responding."²

Here is an everyday example. Ask a colleague what time it is and they answer at once. Ask them to plan a three-city trip on a fixed budget and they get out a notepad first. A reasoning model takes out the notepad. You pay for that notepad, too. OpenAI says the hidden steps "are billed as output tokens," even though you can't see them through its API.¹ So when a reasoning model gives you a better answer, it often costs more and takes longer.

Where the word comes from

We could not find a source that shows who first used the phrase "reasoning model," so we won't name one. The idea underneath it is older than the label. In January 2022, Google researchers showed that a large language model does better on hard problems when it writes out "a series of intermediate reasoning steps." They called this a chain of thought.³ In 2024, OpenAI described its o1 models as "trained with large-scale reinforcement learning to reason using chain of thought."⁴ Reinforcement learning means training by trial, with a reward when the model gets something right. In January 2025, the Chinese lab DeepSeek reported that reasoning abilities "can be incentivized through pure reinforcement learning."⁵ By April 2025, Anthropic wrote that "Since late last year, 'reasoning models' have been everywhere."⁶ The phrase first meant a way of prompting a model. It now means a type of model that is trained to reason this way every time.

How others define it

• OpenAI: "Reasoning models use internal reasoning tokens before producing a response."¹
• Anthropic: "These are AI models—such as Claude 3.7 Sonnet—that show their working."⁶
• Apple researchers: "Large Reasoning Models (LRMs) that generate detailed thinking processes before providing answers."⁷
• Google: "When you use a thinking model, Gemini reasons internally before responding."²

Where they agree and differ

They all agree that a reasoning model thinks before it answers. They differ on two points.

The first is whether you can see the work. Anthropic stresses that these models "show their working."⁶ OpenAI keeps the raw steps hidden from its API customers and shows optional summaries instead.⁸

The second is whether the visible steps are honest, and whether "reasoning" is even the right word. This is the strongest view against taking the label at face value. In Anthropic's tests, when a model used a hint that researchers had planted, it often didn't say so: "Claude 3.7 Sonnet mentioned the hint 25% of the time, and DeepSeek R1 mentioned it 39% of the time."⁶ Apple researchers found that these models "face a complete accuracy collapse beyond certain complexities." They also found that the models' reasoning effort rises with difficulty, "then declines despite having remaining token budget."⁷ Neither study says the models are useless, and both measured improvements in some areas. Still, they are good reasons not to treat the written-out steps as a true record of how the model reached its answer.

In the news

On October 26, 2026, the security firm Xint reported that it had "discovered a bug in the OpenAI API that leaks raw reasoning traces from frontier OpenAI models." A reasoning trace is the model's hidden working. Xint says that "OpenAI confirmed the issue and deployed an undisclosed fix on October 23, 2026."⁸ We are fairly sure the bug was real, but this account comes from the researchers. We have not seen a statement from OpenAI. Xint also says it "can't prove token-for-token fidelity without the original plaintext," meaning it can't prove that what it pulled out matches the hidden reasoning word for word.⁸ The story shows that the hidden working is now valuable enough that people try to steal it. According to Xint, one reason OpenAI hides it is "to prevent distillation," which means using one model's outputs to train another model.⁸

What it means for your organization

The AI agents your teams use more and more often run on reasoning models. That changes three things. Costs grow with how hard the model thinks, not just with how long its answer is.¹ Quality goes up on hard tasks but can fail all at once past a certain level of difficulty.⁷ And the reasoning you see may be a summary, or it may leave things out.⁶⁸ So don't treat a model's shown reasoning as an audit trail. Keep a person responsible for checking the result itself (see [a person in the middle](/definitions/a-person-in-the-middle) and [human in the loop](/definitions/human-in-the-loop)).

One question to ask your team: For the decisions where we use a reasoning model, how do we check the answer without relying on the model's own account of how it got there?