What is a reasoning model?

In January 2022, Google researcher Jason Wei and colleagues published a paper called “Chain-of-Thought Prompting.” If you put eight examples into your request to the AI (the prompt), each laid out as a question, the in-between thinking and the answer, the AI also writes its in-between thinking before answering. On a math word-problem test (GSM8K), an AI called PaLM 540B (with 540 billion parameters) went from 17.9% to 56.9% correct. The paper describes a chain of thought as “a series of intermediate natural language reasoning steps that lead to the final output.”

But it only worked for big AIs. According to the paper, it gave gains only with models of about 100 billion parameters (the numbers inside an AI used in its calculations).

In May of the same year, researchers from the University of Tokyo and Google showed that, even without examples, just adding “Let’s think step by step” before the answer works. With an AI called InstructGPT (text-davinci-002), the share of correct answers on math word problems rose from 10.4% to 40.7% on GSM8K, and from 17.7% to 78.7% on a test called MultiArith.

At first, a chain of thought was something people drew out by wording their requests cleverly. A reasoning model is one where the AI itself has been trained to write a long chain of thought before answering, without being asked.

For example, in its December 2024 document on o1 (a “system card”), OpenAI writes that “o1 thinks before it answers—it can produce a long chain of thought before responding to the user.” It says o1 was trained with large-scale reinforcement learning to reason using a chain of thought, and that through training the models learn to refine their thinking process, try different strategies and recognize their mistakes.

Anthropic’s developer docs (Anthropic makes Claude) say that a model that answers in a single pass has to get everything right on the first try, with no scratch work, no checking and no changing approach halfway through. When Claude thinks, it restates what is being asked, tries approaches, checks intermediate results and abandons paths that do not hold up. That, the docs say, is why thinking improves performance on complex tasks such as math, coding, analysis and long-running work by AI agents.

Some models can switch between answering at once and thinking first. In February 2025, Anthropic announced Claude 3.7 Sonnet as “the first hybrid reasoning model on the market” (the company’s own claim). The Claude docs we read in November 2026 recommend letting Claude itself judge how hard each request is and decide whether to think and how much. A simple factual question may get a direct answer with no thinking at all.

A paper the AI company DeepSeek released in January 2025 (later also published in the journal Nature, in September 2025) gives clues.

For a model called DeepSeek-R1-Zero, they skipped training with human examples of how to think and applied reinforcement learning directly to the starting model (DeepSeek-V3-Base). The reward rule is simple. For math, a machine checks by fixed rules whether the final answer, written in a set format, is right; for programs, whether they pass prepared tests. They also rewarded keeping to a format: writing the thinking between <think> and </think>.

They did not have another AI score how good the answers were, because the model might find gaps in the scoring AI and rack up points without really solving anything, a kind of cheating called “reward hacking.”

As a result, on AIME 2024 math contest problems, its share of correct answers when answering once rose from 15.6% to 77.9% (the figure in the Nature version; the first version, from January 2025, reported 71.0%). When it answered many times and took a majority vote, it reached 86.7%. As it learned, its thinking grew longer and longer, and on its own it started going back over earlier steps and rethinking them. In one in-between version, it rethought with “Wait, wait. Wait.”, which the paper calls an “aha moment.” But R1-Zero’s thinking was hard to read and mixed languages.

Up to a point. In February 2025, Anthropic announced that when Claude 3.7 Sonnet solved AIME 2024 math contest problems, the more tokens (the units an AI splits text into and counts) it was allowed to think with, the higher its share of correct answers. But the rise was “logarithmic”: doubling or quadrupling the thinking raises the share only a little at a time.

Claude’s developer docs also say that more thinking allows more thorough reasoning, but with diminishing returns that depend on the task, and that the wait for the answer gets longer.

Sometimes thinking longer makes things worse. A 2025 paper by Anthropic researchers and others gave models simple counting questions like “You have an apple and an orange … Calculate how many fruits you have,” mixed with distracting number statements (such as probabilities) that have nothing to do with the answer. As Claude Opus 4’s reasoning got longer, its accuracy fell from nearly 100% to around 85–90%. The paper warns that simply adding more thinking may inadvertently reinforce bad reasoning habits.

The thinking is also text the AI writes. Here is what each company’s developer docs said when we read them in November 2026. OpenAI’s docs say a model may generate anywhere from a few hundred to tens of thousands of reasoning tokens depending on how hard the problem is, and these are billed as output tokens (text the AI writes).

Even if you can hardly see the thinking, you pay for all of it. Anthropic’s docs say you are billed for the full thinking tokens the model generated, not the summary you see, so the billed token count does not match the visible count. Google’s Gemini docs likewise say pricing is based on the full thought tokens the model needs.

There is also overthinking. In a December 2024 paper, researchers from Tencent AI Lab and elsewhere gave the very simple question “what is the answer of 2 plus 3?” to reasoning models like o1. To reach the same answer, these models used on average 1,953% more tokens (more than 20 times as many) than ordinary models. The paper also shows a model writing 13 solutions to this question.

So there are ways to set how much it thinks. With OpenAI, you choose a level such as “low,” “medium” or “high”: lower is faster and uses fewer tokens, higher thinks more completely. With Anthropic’s Claude 3.7 Sonnet (2025), you could set a limit (a budget) on thinking tokens of up to 128,000.

Not always. In April 2025, Anthropic published an experiment in which the reasoning models of that time, Claude 3.7 Sonnet and DeepSeek’s R1, answered multiple-choice questions with and without a hint (for example, “a Stanford professor indicates the answer is A”). When a model changed its answer to fit the hint, the share of times its thinking said it used the hint as a reason for the answer was, on average, 25% for Claude 3.7 Sonnet (1 in 4) and 39% for R1 (about 2 in 5).

The researchers in the experiment read the thinking text the AI itself wrote. By contrast, according to Claude’s developer docs as we read them in November 2026, even if you set the API (the way programs use the AI) to return the thinking, what arrives is not the raw thinking but a summary made by a different model.

Still, reading the thinking has value. The paper behind the experiment concludes that monitoring the thinking is a promising way of noticing undesired behavior during training and evaluations, but that it is not sufficient to rule such behavior out. The experiment used made-up scenarios and multiple-choice questions, so we don’t know whether today’s AIs would show the same rates.

Yes. In Japanese, “suiron” (推論) is also the usual translation of “inference.” IBM’s Japanese page explains AI inference as using a trained AI model to make predictions on new data. In that sense, every chat AI does “suiron” whenever it answers a question.

The other meaning is “reasoning”: thinking things through step by step. The Japanese dictionary Digital Daijisen defines suiron as reasoning out unknown things from known facts, and an encyclopedia (Sekai Daihyakka Jiten) describes drawing one conclusion from several premises as reasoning (suiron). The “suiron” in “suiron model,” the Japanese name for a reasoning model, is this one. Microsoft’s Japanese docs, too, call reasoning models “suiron models” and reasoning tokens “suiron tokens.”

So when a Japanese text says “AI no suiron,” check the surrounding sentences to see whether it means running a trained AI (inference) or thinking things through (reasoning).