An AI trained to work through a problem step by step before it answers.
A reasoning model is a language model trained to think before it speaks: it writes out a long series of working steps, often hidden, before giving its answer. It does especially well on multi-step problems in math, science and coding.
OpenAI's o1 (September 2024) made the idea widely known: on a US high-school math contest, OpenAI reported about 74% for o1 versus about 12% for a model that answered directly. In January 2025, DeepSeek-R1 shared a way to build this ability with reinforcement learning. The cost is slower answers and more tokens.
Don't confuse it with inference. Inference simply means running any trained model.
Think of it like this
One student writes the answer straight down; another works it out on scrap paper first. A reasoning model has the scrap-paper habit built in. It solves harder problems, but takes longer.
For example
Give it a tough math puzzle and you may see a "thinking…" note for a while before the answer appears.
Common mix-ups
Mix-upA reasoning model gives faster, better answers to every question.
ActuallyFor questions with a one-step answer, like a country's capital or a spelling check, long thinking adds almost nothing except waiting time. That's why many apps let you choose between a "quick answer" mode and a "think deeply" mode.
Mix-up"Reasoning" and "inference" are the same thing.
ActuallyThey mean different things, but they're linked. While a reasoning model writes out its long working, the model is being run for that much longer, so "thinking more" really means spending more computing on running it.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- Which questions call for a reasoning model, and which for a model that answers straight away?
- Activity
- Make ten question cards (the weather, a word's meaning, a three-step maths word problem, a quick hello…) and have groups sort them into "needs scratch paper" and "fine to answer straight away". Explain your rule, then check the o1 (E0430) and DeepSeek-R1 (E0500) records to see when and how this approach became known.
Records about this
Source-checked moments from AI history.
Proved that reasoning ability is not a secret recipe of closed labs; self-checking, re-examination and strategy switching emerged. The same report also published the RLVR recipe and a recipe for transferring the capabili
The first commercial reasoning model; on AIME 2024, from 12% for GPT-4o to 74% for o1