reasoning models

A reasoning model is one that has been trained to think before it speaks, producing a long stretch of internal working and only then a concise answer. With ordinary models you have to coax step-by-step thinking out with prompts; with a reasoning model it is baked in, and the thinking can run for thousands of tokens, exploring, backtracking, and double-checking before it settles. The visible reply is usually short, but a large hidden draft did the heavy lifting to get there.

These models are typically made by taking a capable base model and training it, often with reinforcement learning, to be rewarded when its long deliberations lead to correct answers on problems with checkable solutions, like maths and code. Over many rounds the model learns habits that pay off: laying out cases, testing its own claims, noticing a contradiction and revising. The trade-off is plain on the receipt and the clock: each answer costs far more compute and waiting time, so reasoning models earn their keep on genuinely hard problems and are overkill for simple ones.