reasoning model
A large language model trained, usually with reinforcement learning, to generate a long series of intermediate steps before its final answer, spending extra computation at answer time.
- OpenAI’s o1, introduced in September 2024
- Google’s Gemini 2.0 Flash Thinking, an experimental model from December 2024
- DeepSeek-R1, an open-weight model trained with reinforcement learning
- Claude 3.7 Sonnet with extended thinking switched on
- Current meaning
- 2024
- OpenAI’s o1, trained by reinforcement learning to think before answering.
- Also called
- Thinking model
- Google’s term. Research papers also say large reasoning model (LRM).
- Opposite
- Standard LLM
- A model that answers without a long reasoning phase, as in Apple’s tests.
In plain words
In AI, a reasoning model is a language model trained to work through a problem step by step before it answers, trading time and cost for accuracy.
Most large language models begin an answer as soon as they read the question. A reasoning model first writes a long stretch of working. It breaks the problem into smaller steps, notices its mistakes and switches approach when one fails. OpenAI introduced the current sense with o1 in September 2024. It said o1 learned these habits in training and did better when given more time to think.1
Google, DeepSeek and Anthropic released their own within six months. DeepSeek’s R1 is an open-weight model, whose weights anyone can download. By 2026 reasoning had become a setting within the main models. Google’s Gemini models, for example, decide by default how much to think about each request.2
OpenAI’s launch results for o1 came from coding contests, a math olympiad qualifier and PhD-level science questions. The costs are time and money. The working is made of tokens, and providers bill it as output even when they hide it.3
The term is often confused with chain of thought, the step-by-step text itself. Researchers first drew that text out of ordinary models by prompting them. A reasoning model is trained to produce it, at length, before it answers.
Labs call the working “thinking”, but it is generated text and may not match how the model reached its answer. In Anthropic’s tests, models that used a planted hint usually left it out of their reasoning.4 Apple researchers reported that such models failed completely on puzzles beyond a certain complexity. A reply by A. Lawsen blamed the test design.5
How it happens
Training starts from a model that has finished pretraining. Developers give it problems a program can mark, such as math questions with one correct result or coding tasks with test cases. The model writes a long chain of steps, then an answer. Reinforcement learning rewards the chains that end in a correct answer, which strengthens the habits that led there. DeepSeek scored answers with fixed rules and avoided a second model trained to judge them.6 In reinforcement learning from human feedback, by contrast, the reward reflects which answers people prefer.
The trained model spends extra computation at inference, the stage when a finished model answers. This is often called test-time compute, and more of it tends to raise scores.1 In December 2024, OpenAI’s o3 scored 75.7% on the ARC-AGI benchmark within a $10,000 compute limit. A version using 172 times the compute scored 87.5%.
Developers can set the amount. OpenAI offers effort levels.3 Anthropic let developers cap Claude 3.7 Sonnet’s thinking with a token budget. Gemini models choose their own amount by default.2 What users see varies too. OpenAI’s API hides the working, and Google shows only the final answer unless summaries are requested.
Origin and first known use
Reasoning, first recorded in the late 1300s, is the verbal noun of reason, which came into English through Old French from Latin ratio, “reckoning, understanding”. The compound took its current AI sense in September 2024. That month OpenAI introduced o1, trained with reinforcement learning to work through a chain of thought, and called the cheaper o1-mini a reasoning model. DeepSeek adopted the term for its R1 models in January 2025. Research papers also say large reasoning model (LRM). Google calls such models thinking models. Anthropic called Claude 3.7 Sonnet a hybrid reasoning model: the same model can answer at once or think step by step.
- 1300s Reasoning, “act or process of thinking logically”, enters English in the late 1300s. It is formed from the verb reason, which goes back through Old French to Latin ratio.Online Etymology Dictionary, “reasoning” and “reason” Etymonline →: source for 1300s
- 2022 Wei and colleagues show that prompting large models with worked examples that spell out intermediate steps improves arithmetic, commonsense and symbolic reasoning. They call it chain of thought prompting.Wei et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” arXiv 2201.11903 →: source for 2022
- 2024 OpenAI introduces o1, trained with reinforcement learning to use a chain of thought. Results cover coding contests, a math olympiad qualifier and PhD-level science. The current sense starts here.OpenAI, “Learning to reason with LLMs”, 12 September 2024 openai.com →: source for 2024
- 2024 OpenAI’s o3 scores 75.7% on ARC-AGI’s semi-private test, and 87.5% with 172 times the compute. François Chollet, writing for ARC Prize, says he does not think o3 is AGI yet.François Chollet, “OpenAI o3 Breakthrough High Score on ARC-AGI-Pub”, ARC Prize, 20 December 2024 arcprize.org →: source for 2024
- 2025 DeepSeek publishes R1 and R1-Zero, reasoning models trained with large-scale reinforcement learning, reports results comparable to OpenAI’s o1, and releases the weights.DeepSeek-AI, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning”, arXiv v1 arXiv 2501.12948 →: source for 2025
- 2025 Anthropic calls Claude 3.7 Sonnet a hybrid reasoning model: it answers at once or thinks step by step, and developers can cap its thinking with a token budget.Anthropic, “Claude 3.7 Sonnet and Claude Code”, 24 February 2025 anthropic.com →: source for 2025
- 2025 Apple researchers compare large reasoning models (LRMs) with standard LLMs on puzzles. Standard models do better on easy ones; beyond a certain complexity, both collapse completely.Shojaee et al., “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity” arXiv 2506.06941 →: source for 2025
- 2025 OpenAI’s GPT-5 pairs an efficient model with a deeper reasoning model, GPT-5 thinking, and a router that picks one for each request.OpenAI, “Introducing GPT-5”, 7 August 2025 openai.com →: source for 2025
In the wild
-
“We’re releasing OpenAI o1‑mini, a cost-efficient reasoning model.”
OpenAI, product announcement openai.com →: OpenAI, product announcement, 12 Sep 2024Archived →: copy of OpenAI, product announcement, 12 Sep 2024 -
‘Gemini 2.0 Flash Thinking’ is Google’s first reasoning model”
-
“We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1.”
-
“Today, we’re announcing Claude 3.7 Sonnet, our most intelligent model to date and the first hybrid reasoning model on the market.”
Anthropic, product announcement anthropic.com →: Anthropic, product announcement, 24 Feb 2025Archived →: copy of Anthropic, product announcement, 24 Feb 2025 -
“Since late last year, “reasoning models” have been everywhere.”
-
“Recent generations of language models have introduced Large Reasoning Models (LRMs) that generate detailed thinking processes before providing answers.”
Shojaee et al., Apple, arXiv preprint arXiv 2506.06941 →: Shojaee et al., Apple, arXiv preprint, 7 Jun 2025Archived →: copy of Shojaee et al., Apple, arXiv preprint, 7 Jun 2025 -
“GPT‑5 is a unified system with a smart, efficient model that answers most questions, a deeper reasoning model (GPT‑5 thinking) for harder problems…”
OpenAI, product announcement openai.com →: OpenAI, product announcement, 7 Aug 2025Archived →: copy of OpenAI, product announcement, 7 Aug 2025 -
“When you use a thinking model, Gemini reasons internally before responding.”
Google, Gemini API documentation ai.google.dev →: Google, Gemini API documentation, 23 Sep 2026Archived →: copy of Google, Gemini API documentation, 23 Sep 2026
Not to be confused with
-
Chain of thought
Chain of thought is the step-by-step text on the way to an answer, first drawn out of ordinary models by prompting. A reasoning model is trained, usually with reinforcement learning, to write long chains before answering.
-
Artificial general intelligence
AGI usually means a hypothetical system with broadly human-level ability across tasks. Reasoning models are language models that do better on hard math, science and coding tests. Writing for ARC Prize, François Chollet judged o3 not yet AGI.
-
Human thinking
Labs call the working “thinking”, but it is generated text and not a reliable record of how the answer was reached. In Anthropic’s tests, models often left out a hint that had shaped their answer.
Questions
What is a reasoning model in AI?
A reasoning model is a large language model trained, usually with reinforcement learning, to write out a long series of intermediate steps before its final answer. OpenAI introduced the current sense with o1 in September 2024. Google, DeepSeek and Anthropic released their own within six months. OpenAI’s launch results for o1 came from coding contests, a math olympiad qualifier and PhD-level science questions.
What is the difference between a reasoning model and chain of thought?
Chain of thought is the step-by-step text a model writes on the way to an answer. In 2022, Jason Wei and colleagues drew it out of ordinary large models by prompting them with worked examples. A reasoning model is trained to produce long chains of thought before it answers, usually through reinforcement learning that rewards chains ending in a correct answer.
Why are reasoning models slower and more expensive?
Reasoning models generate a long stretch of working, made of tokens, before the final answer, and each token takes time and computation. Providers bill that working as output. OpenAI’s API bills reasoning tokens even though it does not show them, and Google bills Gemini’s thinking tokens. Developers can usually set how much a model thinks, trading speed and cost against accuracy.
Do reasoning models really think?
Labs call the working “thinking”, but it is generated text and may not show how the model reached its answer. In 2025 Anthropic planted hints in test questions. When Claude 3.7 Sonnet and DeepSeek R1 used a hint, their reasoning mentioned it 25% and 39% of the time respectively. Apple researchers also reported collapse on hard puzzles, though a reply by A. Lawsen blamed the test design.
Is a reasoning model the same as AGI?
No. Artificial general intelligence usually means a hypothetical system with broadly human-level ability across tasks. Reasoning models are language models trained to work through problems at length. In December 2024, a high-compute version of OpenAI’s o3 scored 87.5% on a semi-private ARC-AGI test. François Chollet, writing for ARC Prize, still said he did not think o3 was AGI yet.
Watch
- Adv. LLM Agents MOOC | UC Berkeley Sp25 | Inference-Time Techniques for LLM Reasoning by Xinyun ChenA Google DeepMind researcher’s UC Berkeley lecture on spending more computation at answer time to improve reasoning.
- Adv. LLM Agents MOOC | UC Berkeley CS294-280 Sp25 | Learning to Reason with LLMs by Jason WestonMeta’s Jason Weston lectures at UC Berkeley on methods for training language models to reason.
- Adv. LLM Agents MOOC | UC Berkeley Sp25 | Open Training Recipes: LLM Reasoning by Hanna HajishirziHanna Hajishirzi of the University of Washington on openly published recipes for training language models to reason.
Cite this entry
“Reasoning model.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/
Superrduperr. (2026, September 24). Reasoning model. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “Reasoning model.” SUPERR_dictionary. September 24, 2026. https://superrduperr.com/