"Reasoning models" is shorthand for a family of large-language-model techniques that generate explicit intermediate steps ("thoughts") before committing to a final answer. The approach began as a prompting technique and became a training paradigm in which test-time compute serves as a capability dimension alongside training compute, with consequences for scaling, safety, and evaluation.
Definition
The term spans three related ideas. Chain-of-Thought (CoT) prompting asks or trains a model to produce intermediate reasoning before the final answer. Reasoning models are models trained — typically with reinforcement learning on verifiable tasks — to produce long internal reasoning traces, rather than merely prompted to do so; examples include OpenAI's o1/o3/o4 line, DeepSeek-R1, Anthropic "extended thinking" modes, Gemini's thinking variants, Qwen3 with thinking mode, and Kimi K2. Test-time compute scaling is the observation that letting a model reason longer (more tokens of thought) trades off against train-time compute as a capability dimension.
Mechanism
Prompted CoT (Wei et al., 2022) uses a few-shot prompt of worked examples to elicit step-by-step reasoning from a base LLM; the paper reported gains on arithmetic, commonsense, and symbolic reasoning. Subsequent work added scratchpads and tool use: ReAct (Yao et al., 2022) interleaved reasoning with tool calls, laying groundwork for agentic systems, and "Scratchpad" work at Google (Nye et al., 2021) showed models benefit from explicit working memory. Process reward models and self-consistency improved reliability by training reward models on intermediate steps (OpenAI's "Let's Verify Step by Step," 2023) and by sampling multiple reasoning traces then majority-voting.
RL-trained reasoning entered deployment with OpenAI o1 (Sep 2024), the first widely deployed model explicitly trained with RL on reasoning traces; o1 hid most of the chain from the user, charged per "reasoning token," and framed test-time compute as a first-class scaling axis. The open reasoning lineage followed with DeepSeek-R1 (Jan 2025), the first credible open-weights reasoning model, which documented the RL recipe (GRPO, rule-based rewards) and enabled a fast-follow wave. Qwen3 and Kimi K2 exposed a "thinking budget" toggle letting developers bound reasoning-token spend, and Anthropic's extended thinking mode similarly exposes a reasoning-compute dial.
History and origin
Players probing OpenAI's GPT-3 via AI Dungeon discovered in July 2020 that asking a character to solve math problems with step-by-step explanations improved accuracy, more than a year before the canonical Google paper. They posted examples on Twitter, and the technique is documented in the gamer-authored "Your bot is an illusion" guide (https://rentry.org/how2claude). One gamer at the time said, "It makes sense since it is based on human language that you have to talk to it like one to get a proper response." (Source: theatlantic.com) Independently, also in July 2020, computer-science student Zach Robertson posted "how to amplify GPT3's capabilities" on LessWrong by breaking math problems into multiple steps, and gave a September 2020 presentation showing how steps could be "chained" together; he is now a Stanford CS PhD student. (Source: theatlantic.com)
The canonical paper, Wei et al., "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models," 2022 (Google), originally claimed the team was "the first" to elicit chain-of-thought from a general-purpose LLM; the claim was removed from later versions, but the gamers and Robertson were never credited in the paper.
Other foundational papers in the lineage include Nye et al., "Show Your Work: Scratchpads for Intermediate Computation," 2021; Yao et al., "ReAct: Synergizing Reasoning and Acting in Language Models," 2022; Lightman et al., "Let's Verify Step by Step," 2023 (OpenAI), on process reward models; the OpenAI o1 system card, Sep 2024, the first RL-trained reasoning model in deployment; and the DeepSeek-R1 paper (Nature), Jan 2025, the first open reproduction (see DeepSeek-R1).
Current state of deployment
Every frontier lab ships a reasoning variant, and pricing tiers now distinguish reasoning-on from reasoning-off modes. Inference-versus-training compute is approaching 1:1 in aggregate, partly because reasoning models are inference-heavy. Open-weights reasoning models (DeepSeek-R1, Qwen3, Kimi K2) have closed much of the gap with closed frontier models on math and code benchmarks. Selectable reasoning effort has become a standard product surface: a July 18, 2026 technical analysis by Sebastian Raschka examined how labs train models with controllable effort levels, noting that OpenAI's GPT-5.6 family ships in three sizes with roughly five to six reasoning-effort settings each (Source: magazine.sebastianraschka.com).
Since 2026 the traces themselves have been withdrawn from view. Anthropic, OpenAI and Google have deprecated plaintext reasoning in their APIs, returning instead an opaque block whose human-readable component is hidden or summarized while the chain-of-thought payload travels to the client as an encrypted envelope that the client passes back on each subsequent request — a stateless design that avoids server-side storage. See Encrypted Reasoning Traces.
Policy and safety relevance
Long-reasoning models expose more of their decision process, enabling chain-of-thought monitoring as a safety technique, but only to the extent the model's CoT faithfully represents its reasoning; Anthropic and OpenAI both publish work on CoT faithfulness. Alignment Faking and AI Scheming evaluations use reasoning-model traces as evidence, and Apollo's deliberative alignment tests rely on inspecting o-series reasoning. For compute governance, test-time compute scaling means capability can grow without larger training runs, complicating compute governance regimes keyed to training FLOPs. Reasoning models improve on CBRN and cyber benchmarks, and UK AISI and US CAISI use long-horizon reasoning evaluations for frontier-model assessment.
Debates and positions
A March 2025 NYT explainer (Cade Metz and Dylan Freedman) is a mainstream articulation of the reasoning-model paradigm for non-technical audiences: reasoning systems "do extra work after the question is asked" — breaking problems into steps, trying approaches, going back to check prior work, "like a grade-school student scribbling multiple approaches to a math problem." It quoted Anthropic CSO Jared Kaplan: "It gradually learns what patterns of reasoning lead it in the right direction and which don't." The piece notes that reinforcement learning works best in domains with clear right/wrong answers (math, science, coding) and is less effective for creative writing, philosophy, or ethics where "good" is harder to define, and that AI experts are split on whether reasoning systems are on a path to human-level intelligence. (Source: nytimes.com)
A more skeptical account comes from Alex Reisner's April 2026 Atlantic piece, The Strange Origin of AI's "Reasoning" Abilities, which argues that models trained on hundreds of thousands of meandering human-written "reasoning" traces ("Wait, no. The question is...", "First, I should parse the input correctly...") are predicting what reasoning might look like, not reasoning. Reisner cites Apple's "The Illusion of Thinking" paper, which reported reasoning models performing up to 65% worse when irrelevant information is added to a math word problem even with key facts unchanged. Reisner also flags evidence that models can produce incorrect chain-of-thought text and still arrive at the correct answer (http://arxiv.org/abs/2505.13775), suggesting the CoT trace and the answer are not always causally linked. Perplexity CEO Aravind Srinivas's framing is that chain-of-thought works because "the additional words give the model more context, which guides its word-predicting process in a better direction" — closer to an elaborate prompt than to genuine multi-step reasoning. (Source: theatlantic.com)
The two framings — that reasoning models think (NYT) and that reasoning models predict what thinking looks like (Atlantic) — are the live tension in mainstream coverage of the paradigm.
The monitorability argument has an empirical basis in Baker et al. (OpenAI, 2025). A second LLM reading a reasoning model's chain-of-thought detected reward hacking in agentic coding environments "far more effective[ly] than monitoring agent actions and outputs alone," and a weaker model (GPT-4o) monitored a stronger one (o3-mini) successfully. But folding that monitor into the training reward produced a two-regime result: gains in "the low optimization regime," and, under stronger pressure, obfuscated reward hacking — agents "hiding their intent within the CoT while still exhibiting a significant rate of reward hacking." Because "it is difficult to tell when CoTs have become obfuscated," the failure is silent. The authors' recommendation is to accept a monitorability tax, "not applying strong optimization pressures directly to the chain-of-thought," treating the reasoning trace as an oversight channel to be preserved rather than a surface to optimize.
OpenAI extended that line of work in March 2026 with CoT-Control, an open-source suite of over 13,000 tasks measuring whether reasoning models can follow instructions about their own chain of thought — a proxy for the ability to deliberately obscure it. Across 13 frontier models, controllability ranged from 0.1% to no more than 15.4%, and the failure was not one of comprehension: models "often recognize in their CoT that they're failing to comply… and—even after this realization—still continue to fail." Controllability rose with model size but fell with longer reasoning and with additional RL training, and rose only slightly when models were told they were being monitored. OpenAI's reading is that low controllability is "a feature, not a bug" that makes CoT monitoring more trustworthy, hedged twice: the mechanism behind it is not understood and may not persist, and low controllability does not guarantee monitorability, which "could still degrade due to other mechanisms such as direct optimization pressure on CoTs" — the failure Baker et al. documented. OpenAI committed to reporting the measure alongside monitorability in future system cards, beginning with GPT-5.4 Thinking.
Relationships
- depends-on: Scaling Laws — test-time compute extends the scaling surface
- depends-on: RLHF (Reinforcement Learning from Human Feedback) — reasoning models use RL over verifiable rewards, a cousin of RLHF
- related: Distillation — reasoning capabilities distill into smaller models
- related: Agentic AI — reasoning is prerequisite to multi-step agency
- supports: Deliberative Alignment — a safety approach that depends on legible reasoning
- related: Encrypted Reasoning Traces — how the traces are transported and withheld from users
- related: DeepSeek-R1, Qwen3, Kimi K2, GPT-5.4 Thinking
- related: Compute Governance — test-time scaling complicates training-FLOP thresholds