What is a reasoning model?
A reasoning model is a large language model (LLM) that has been trained to generate a chain of intermediate steps before it commits to an answer. Ordinary chat models produce the answer directly, one token after the next. A reasoning model first writes out a working section, sometimes called thinking or a reasoning trace, in which it breaks the problem down, tries approaches, checks its own arithmetic, and discards wrong turns. Only then does it write the response the user sees.
The behavior is related to chain-of-thought prompting, where a user asks a standard model to think step by step. The difference is that a reasoning model has been trained, usually with reinforcement learning on problems with checkable answers, to do this on its own and to do it well. Many vendors expose a setting that controls how much effort the model spends, from a short pass to a long deliberation.
Why a reasoning model matters
Reasoning models do better on tasks with many dependent steps: multi-part calculations, planning, debugging code, interpreting a contract clause against several definitions, or reconciling figures across documents. They are also better at noticing when a question is underspecified. The cost is time and tokens. The reasoning trace can run to thousands of tokens that the customer pays for under token-based pricing and that add seconds or minutes to time to first token. For a quick lookup or a simple rewrite, a standard model is the better choice.
Reasoning model in practice
A building products distributor uses an assistant to review customer pricing exceptions. Each case involves a contract price list, a rebate schedule, freight terms, and a quote that a salesperson wants to honor. A standard model tended to pick one rule and ignore the others. A reasoning model works through the rules in order, shows which one governs, and flags cases where the quote cannot be reconciled with any of them. The review takes longer per case, so the team runs it only on exceptions above a set margin threshold and keeps the standard model for routine quotes.
What to watch for
- Match the model to the task, because routing easy requests to a reasoning model wastes money and makes users wait.
- Confirm whether the reasoning trace is billed, whether you can see it, and whether it is logged, since it may contain sensitive data from the prompt.
- Test with your own hard cases, because a reasoning model can still hallucinate a confident wrong answer after a long chain of steps.
- Check that your timeouts and latency targets allow for long generations before you put one behind a chat interface.