A reasoning model is a large language model trained to work through a problem in a long series of intermediate steps before it gives its final answer. It spends more time and computing power on each question, and in return does markedly better on mathematics, programming, science and planning.
From prompting trick to training goal
Researchers noticed in 2022 that models answered harder questions more reliably when prompted to write out their steps, a technique known as chain of thought. Reasoning models turn that habit into a trained skill. OpenAI’s o1, previewed in September 2024, was the first widely used example: it was trained with reinforcement learning to produce long internal reasoning, and OpenAI reported that both extra training and extra thinking time raised its scores. In January 2025, the Chinese lab DeepSeek released R1, an open-weights reasoning model, and described how reinforcement learning on problems with checkable answers – math results, code that passes tests – had taught an experimental version, R1-Zero, to reflect, check itself and try other approaches without being shown how. After o1, Google, Anthropic and others released their own reasoning models, often “hybrid” models that can answer quickly or think at length, and users or developers can set how much thinking to allow.
What changes
Reasoning models move part of the effort from training to the moment of use, a shift described under test-time compute. They generate many more tokens per answer, which makes them slower and more expensive per question; some providers show the reasoning, others show only a summary or nothing. The gains have been large on hard benchmarks: in July 2025, experimental reasoning models from Google DeepMind and OpenAI reached gold-medal scores (35 of 42 points) at the International Mathematical Olympiad, writing their proofs in natural language within the contest’s time limits; DeepMind’s result was certified by the organizers, OpenAI’s was graded by former medalists. Their reasoning also gives safety researchers a window into how a model arrives at an answer, though not a perfectly reliable one.
Limits and debate
Reasoning models still make mistakes and sometimes produce long, confident reasoning toward a wrong answer. A June 2025 study by Apple researchers, “The Illusion of Thinking,” reported that several reasoning models collapsed on puzzles beyond a certain complexity and, near that point, used fewer reasoning tokens despite having budget left. A widely shared reply argued that some test cases were unsolvable or exceeded the models’ output limits; its author later said parts of it were written half in jest, though he stood by the point about the unsolvable puzzles. Whether such systems “really” reason is partly a question of definition. Few dispute that the method made models markedly better at exactly the tasks – proofs, code, multi-step analysis – where earlier chatbots were weakest, though some researchers argue the training mainly draws out abilities the base model already had, and that it pushed AI further into the work of AI agents.
Why it matters
Reasoning models changed what “a better model” means. Progress no longer comes only from building bigger models but also from letting them think longer, which affects pricing, energy use and how governments try to measure frontier capability (see compute governance). For users, the practical lesson is simple: for a hard question, a model that takes its time is often the better choice – and its visible reasoning is worth reading, not just its conclusion – with the caveat that it may be a summary or incomplete (see chain of thought).
Written with Claude, made by Anthropic, which offers Claude models with an extended-thinking mode; a Claude model was among those tested in the Apple study, and Claude Opus was credited as co-author of the first version of one cited reply.
Related
Chain of Thought · Test-Time Compute · Inference · Large Language Model · Reward Hacking · AI Agent
Sources: OpenAI: Learning to reason with LLMs (September 12, 2024) · DeepSeek-AI: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (arXiv) (January 2025; published in Nature, September 2025) · Anthropic: Claude 3.7 Sonnet and Claude Code (February 24, 2025) · Anthropic: Claude’s extended thinking (February 24, 2025) · Google DeepMind: Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad (July 21, 2025) · Alexander Wei (OpenAI) on X and OpenAI IMO 2025 proofs (GitHub) (July 19, 2025) · Parshin Shojaee et al.: The Illusion of Thinking (Apple Machine Learning Research) (June 2025) · A. Lawsen: Comment on The Illusion of Thinking (arXiv; the first version credited “C. Opus, Anthropic” as co-author) (June 2025) · Alex Lawsen: When your joke paper goes viral (Substack) (June 15, 2025) · Yang Yue et al.: Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? (arXiv) (April 2025) · Toby Ord: Inference Scaling Reshapes AI Governance (arXiv) (2025)
Leave a Reply