Models
Reasoning Model
Also known as: thinking model
A newer type of language model trained to "think" at length, working through a problem internally before giving its answer.
A reasoning model is a newer kind of AI that deliberately takes time to think before answering. Instead of replying instantly, it works through the problem step by step in a kind of scratchpad, then gives you the result. This makes it much better at hard problems, maths, logic, coding, planning, though it is slower and costs a bit more per answer.
A reasoning model is a language model trained to generate an extended internal reasoning process, often hidden or summarised, before producing its final answer. Where an ordinary model gives a quick reply, a reasoning model spends more effort at answer time to improve accuracy on complex, multi-step tasks.
What changed
Earlier, users had to coax step-by-step thinking with chain-of-thought prompts. Reasoning models bake that behaviour in: they are trained (often with reinforcement learning) to think longer when a problem is hard.
The trade-off
More thinking means better results on difficult problems but slower, costlier responses. For simple tasks a standard model is fine; reasoning models earn their keep on maths, science, coding and planning.
Reasoning models internalise long-form deliberation, generating extended intermediate token sequences (a learned chain-of-thought) before emitting an answer, and are typically trained with reinforcement learning against verifiable reward signals (e.g. correctness on maths/code) to elicit and refine that behaviour (cf. DeepSeek-R1, 2025). This operationalises test-time compute scaling: allocating more inference computation to a single problem yields accuracy gains distinct from scaling model size or training data.
Design and evaluation implications follow. The reasoning trace is often withheld or summarised (both to protect it and because it is not a guaranteed-faithful account of internal computation). Cost and latency scale with the length of deliberation, motivating adaptive/controllable "effort" and routing between fast and reasoning modes. Strengths concentrate on tasks with checkable structure (mathematics, formal logic, programming, multi-step planning); gains are smaller on open-ended generation, and extended reasoning does not eliminate hallucination.