A newer class of model that spends longer thinking before answering, sometimes working through a problem at length internally before producing a response. They cost more and take longer, but do markedly better on hard problems in mathematics, coding and logic. For simple questions they are usually overkill.
Technical view
The insight is that quality can be bought with inference-time computation rather than only with training-time scale. These models are typically trained with reinforcement learning on verifiable tasks where correctness can be checked automatically. Techniques include extended internal reasoning, sampling multiple candidate solutions and selecting among them, and iterative self-correction. The trade-off is a substantial rise in latency and token cost, and the gains are concentrated in domains with checkable answers.
Related
Part of AI Terminology Index, section: Controlling and Extending Behaviour at Run Time.