Reasoning model
A language model designed to work through a problem step by step before giving its answer, rather than replying instantly. This deliberate "thinking" makes it markedly better at maths, coding, and logic, at the cost of taking longer.
An ordinary language model answers in one smooth motion, generating its reply without any visible working. That is fine for a quick fact or a turn of phrase, but it struggles with problems that need several careful steps, where a single early mistake ruins the result. Reasoning models are built to slow down. Before committing to an answer, they generate an internal chain of thought — a worked-out sequence of intermediate steps — and only then respond.
The gain is real. On maths problems, programming tasks, and logic puzzles, letting a model reason before answering can lift its accuracy substantially. The models are often trained for this with reinforcement learning, rewarded not just for the final answer but for reaching it soundly. The trade-off is time and cost: all that thinking uses more computation, so a reasoning model is slower and more expensive than a quick reply.
The term has become one of the clearest dividing lines in recent AI, separating fast general-purpose assistants from the more deliberate systems reached for when a problem genuinely needs to be got right.