Reasoning in AI

Reasoning in AI means that a model works through a problem in steps before it answers, the way you would with pen and paper. Reasoning models such as OpenAI's o1 and DeepSeek-R1 first generate those steps and only then give their answer. That helps with problems that take several steps, such as maths or code. The answer takes longer and costs more, because the thinking is billed as text.

How did it start?

It started as a way of prompting. Wei et al. (Google, 2022) put a few worked examples with intermediate steps in the prompt and called this chain-of-thought prompting. With eight such examples, Google's 540-billion-parameter PaLM set a new best score on GSM8K, a set of maths word problems. Kojima et al. (2022) found that a single sentence often does the job: adding "Let's think step by step" raised the accuracy of OpenAI's text-davinci-002 on the MultiArith sums from 17.7% to 78.7%. Tree of Thoughts (Yao et al., 2023) let the model try several lines of reasoning, judge them and backtrack. In the Game of 24, GPT-4 solved 4% of the puzzles with chain of thought and 74% with this method.

What is a reasoning model?

A reasoning model is trained to think before it answers. OpenAI presented o1 on 12 September 2024: reinforcement learning taught it to use a chain of thought, and it did better both with more training and with more time to think while answering. On the 2024 AIME maths exams it solved 74% of the problems with one attempt each, against 12% for GPT-4o. OpenAI chose to hide the raw chain of thought and show a summary. DeepSeek-R1 (January 2025), released as an open model, showed that reinforcement learning can teach reasoning without human-written examples of it, with habits such as checking its own work emerging along the way.

Anthropic added extended thinking with Claude 3.7 Sonnet in February 2025: the model could answer at once or think step by step first, within a budget set by the developer. With Claude Opus 4.6 in February 2026 came adaptive thinking, in which the model decides for itself when deeper reasoning helps. Researchers also borrow a distinction from psychology: fast, intuitive System 1 thinking versus slow, deliberate System 2 thinking, as in the title of a 2025 survey of reasoning models, From System 1 to System 2.

How does it work?

The reasoning is text. The model writes intermediate steps as tokens and bases its final answer on them. According to Anthropic's documentation (September 2026), Claude restates the question, tries approaches, checks intermediate results and drops paths that do not hold up. Those thinking tokens are billed as output tokens, even when you do not see them. What you can see is at most a summary, and on Claude Opus 5.5 and other current models the thinking is hidden unless you ask for it.

Where does it go wrong?

• The written reasoning is not always the real reasoning. When Anthropic researchers slipped models a hint about the answer, Claude 3.7 Sonnet mentioned the hint it used 25% of the time, and DeepSeek-R1 39% of the time (Anthropic, April 2025).
• More thinking has limits. Apple researchers (Shojaee et al., June 2025) saw standard models do better on easy puzzles, reasoning models win on medium ones, and both collapse on hard ones. A reply by Lawsen argued that the collapse mostly reflected the test design, such as output limits and puzzles without a solution.
• Easy questions can get too much thought. Chen et al. (2024) call this overthinking: extra computation on simple problems with little benefit.

Example

Wei et al. show the difference with a sum. The question: a cafeteria had 23 apples, used 20 to make lunch and bought 6 more; how many does it have? With an ordinary example in the prompt, the model answered 27. With an example that spelled out its steps, it first wrote 23 − 20 = 3, then 3 + 6 = 9, and answered 9.