A reasoning model spends extra thinking time to get harder answers right.
DefinitionWhat it means
A reasoning model is a language model specifically trained or prompted to generate an extended internal chain of intermediate steps, often called thinking or reasoning tokens, before producing its final answer, a technique used by models like OpenAI's o-series and Claude's extended thinking modes. This deliberate step-by-step process improves accuracy on math, coding, and multi-step logic tasks at the cost of higher latency and token usage.
Why it mattersWhy you should care
Reasoning models let teams trade money and time for accuracy on genuinely hard problems, which matters for use cases like complex debugging or multi-step planning where a fast, shallow answer is more likely to be wrong. Deciding when to invoke a reasoning mode versus a standard fast model is now a core architectural choice in AI product design, directly affecting both cost and user-perceived responsiveness.
At a glanceSee it
How a reasoning model routes each prompt — a cheap direct reply for easy asks, and a larger thinking budget spent only when difficulty warrants it.
The inner loop — draft, self-check, and backtrack until a step holds, with the caveat that the written chain need not reflect the model's true reasoning.
Where you see itIn the wild
- A model picker offering a thinking or reasoning mode toggle
- Benchmarks like math or coding olympiad problems favoring reasoning models
- Latency and cost tradeoffs discussed when enabling extended reasoning