For simple tasks such as completing a function, disabling or reducing reasoning depth can save more than 80% of token consumption, while complex cross-module refactoring tasks are worth enabling the highest reasoning depth.