It’s the difference between a quick response and a more deliberate pass through the problem. A reasoning setting or reasoning model gives the system more room to work through logic before producing the answer, which can help on harder tasks but costs more time and usage.
Think about an employee sorting mail versus solving a client issue. Sorting mail is fast: look at the name, drop it in the right slot, move on. Solving a client issue takes more thought, comparing options and noticing tradeoffs before deciding. Reasoning is that second kind of work.
How it shows up
Categorizing a QuickBooks transaction doesn’t need much logic if the rules are clear; explaining a complicated concept, designing a workflow, or debugging a subtle issue does. In products you’ll see this as a model choice, a reasoning-effort setting, or a “think longer” mode. The label changes, the tradeoff stays: more reasoning can improve harder work, but it isn’t free. This connects to model (some answer quickly, some are better at multi-step analysis) and to inference, the run where the model turns input into output. Good prompting still matters: if your request is vague, a prompt framework gives the model the context it needs so reasoning spends effort on the real problem.
Why you care
Reasoning doesn’t guarantee truth. A model can think through bad assumptions and still land on a confident wrong answer, which is why hallucination and review still matter. Not every task deserves the same thought: use fast modes for obvious work, deeper reasoning when a shallow answer would cost more. The point isn’t to make every answer slower. It’s to match the thinking budget to the job.