Reasoning effort gives an application a way to trade extra model computation for the possibility of a more carefully worked answer. Lower settings generally aim for faster, cheaper responses, while higher settings may spend more time or tokens on difficult planning and analysis.
The best setting depends on the task. Formatting a familiar result may work well with low effort, while a difficult coding change or multi-step decision may benefit from more. Higher effort is not a guarantee of correctness and can add latency or cost without improving a simple request.
Evaluations should compare both output quality and resource use at each setting. A model that performs useful work at low reasoning effort can be valuable for high-volume applications even if a higher setting remains available for harder cases.
ELI5
Reasoning effort is a control for how much work an AI model does before answering. It is similar to choosing between a quick mental estimate and sitting down to work through a problem step by step.
For example, a low setting might be enough to rename a variable in a small code sample, while a complex bug across many files may need more effort. More effort can take longer and cost more, and it still does not guarantee the right answer.
