Reasoning model

Reasoning models build on chain-of-thought prompting, which showed that asking for intermediate steps improves answers to complex problems. Instead of relying on the prompt, they are trained, often with reinforcement learning, to produce long internal reasoning traces and to check and revise their work. Many let you choose how much reasoning effort to spend; some show a summary of the reasoning while others hide it.

Why it matters: for hard analytical work, a reasoning model can be more accurate than a standard model, but it uses many more output tokens and takes longer. For simple tasks such as rewriting an email, a standard model is usually faster and cheaper. When comparing plans and APIs, check whether reasoning tokens are billed and how usage limits count them.

Example: asked to find the error in a long chain of spreadsheet formulas, a reasoning model may spend several seconds working through each step before answering, where a standard model replies at once.

Limits: visible reasoning is not a guarantee of correctness, and the model can still hallucinate or reason from false premises. Longer thinking also raises cost and latency.