When should I use light, medium, or high reasoning for an AI request?

Choose the lowest reasoning level that can reliably complete the task. More reasoning can improve multi-step analysis, but it usually takes longer and costs more. It does not automatically make every response better.

Light reasoning suits direct transformations and familiar tasks: rewrite a paragraph, extract obvious fields, brainstorm names, format notes, translate straightforward text, or answer a simple question. The task should have a clear path and a low cost of minor error.

Medium reasoning is a practical default for planning, comparison, document analysis, structured writing, and research synthesis. Use it when the model must connect several pieces of information, respect multiple constraints, or explain a recommendation. Most normal knowledge-work requests fit here.

High reasoning is useful for difficult debugging, complex trade-offs, ambiguous evidence, multi-stage calculations, detailed strategy, or tasks where the model should test competing explanations. It is most valuable when the problem genuinely requires deeper analysis and you are willing to wait for it.

Reasoning level does not replace tools or evidence. A high-reasoning model still needs live search for current information, the actual file for document-specific questions, and a calculator or deterministic code for calculations that must be exact. It can reason carefully from incorrect inputs and still reach the wrong answer.

Improve the prompt before increasing the reasoning level. Clarify the goal, add missing context, remove conflicting instructions, and specify the output. If a medium response is wrong because the model lacked a document or current fact, selecting high will not fix the missing evidence. Increase reasoning when the difficulty is analysis, not when the problem is absent information.