What does the 'adaptive thinking' checkbox in Claude do, and should I leave it on?

Adaptive thinking is a setting in Claude's chat interface that, when checked, significantly increases token usage — reportedly around quadrupling it. Unchecking it makes Claude only slightly 'dumber' but not enough to matter for most work, and it conserves tokens substantially. The recommendation was to leave it unchecked by default and use Sonnet (the default model) without adaptive thinking, only enabling it if you're prepared to burn extra tokens and feel the default model isn't performing well enough — and even then, trying Opus without adaptive thinking first was suggested as a better tradeoff.