Definition
Reasoning models, starting with OpenAI's o1 in September 2024, are trained with reinforcement learning to produce a long internal chain of thought before their final answer. Giving them more 'thinking' time at inference generally improves results on maths, coding, planning and analysis.
The trade-off is cost and latency: a reasoning model may use many times more tokens than a standard model for the same question. Many providers now let developers set how much the model should think, so teams can save the extra effort for the steps that need it.