Standard models answer at a fixed cost. Reasoning models (OpenAI's o-series, Claude's extended thinking, DeepSeek-R1) are trained with reinforcement learning to produce a long private chain of thought first — exploring, backtracking, checking — then answer. On competition math and hard debugging, the gains are dramatic.

The cost is real: more tokens, more latency, more spend per query — and on easy questions, mostly wasted. Matching the thinking budget to the problem is the new tuning knob, which is what "test-time compute" discussions are about.