Reasoning models are an innovation in which a model 'thinks' more—engaging an extended chain of thought and spending more compute at inference time before producing an output—rather than simply producing a single output, which shifts cost from expensive training toward longer runtime.