Definition
A reasoning model is a model trained and served to spend additional inference effort on difficult, multi-step problems before producing an answer or action. The model may plan, compare alternatives, revise intermediate work, or use tools during that process. Providers often expose controls such as reasoning effort or a thinking budget to trade cost and latency against task performance.
The category is functional, not a claim about consciousness. It also does not mean that every output includes valid reasoning. A reasoning model can make factual errors, optimize the wrong objective, or produce a confident answer from a flawed intermediate path.
Origin and usage
Models that perform multi-step inference predate the current label, and the available primary record does not establish one coiner. OpenAI's September 2024 o1 release made "reasoning model" a widely used product category. OpenAI described o1 as trained with reinforcement learning to refine its chain of thought and reported gains from additional test-time compute. That release marks an influential family, not the invention of machine reasoning.
Operational significance
Reasoning models are often useful when a task has hard constraints, uncertain intermediate steps, code or mathematics, or a need to inspect evidence before acting. They can be wasteful for classification, routing, extraction, or other narrow tasks that a faster model handles reliably. Route by measured task performance, latency, and total token use rather than by the prestige of a model label.
Some providers do not expose the model's raw intermediate reasoning. Evaluate the final result, tool trace, cited evidence, and externally checkable work instead of treating hidden reasoning as an audit log.
Distinguish it from nearby terms
- Chain-of-thought prompting asks a model to produce or use intermediate steps. A reasoning model is trained or configured to allocate inference work even without that exact prompt.
- Test-time compute is the broader amount of computation spent at inference. Reasoning models are one way to use more of it.
- A reasoning token is a provider accounting unit for some intermediate generation. It is not a unit of correctness or insight.
Check your understanding
A routing system sends every request to its most expensive reasoning model. Identify two task types that may not benefit and the measurements needed to justify a cheaper route.