Reasoning Model

A reasoning model is a class of large language model that generates an internal chain of thought — sometimes called a "scratchpad" — before delivering its final response. Rather than producing an answer in one pass, the model works through sub-problems step by step, checking its own logic along the way.

This approach dramatically improves performance on tasks that require multi-step math, logical deduction, code generation, and scientific reasoning. OpenAI's o-series models and Anthropic's extended thinking mode are prominent examples. Reasoning models typically trade higher latency and token cost for significantly better accuracy on hard problems.

In enterprise AI deployments, reasoning models are preferred for tasks like financial analysis, legal document review, and complex workflow planning where correctness outweighs speed.