Techniques & Methods
Reasoning Effort in plain English.
Also known as: thinking budget,adaptive thinking,effort level
The one-sentence version
A setting that controls how much hidden thinking a reasoning model does before answering, trading speed and cost against accuracy.
Reasoning effort is the dial that tells a reasoning model how hard to think before it replies. Models like GPT-5, Claude Opus, and Gemini can spend anywhere from a few hundred to tens of thousands of hidden tokens working through a problem; more thinking usually means better answers on hard tasks but slower responses and higher bills, and it adds nothing on easy ones. Vendors expose this as an effort level (low, medium, high), a token budget, or an adaptive mode where the model decides for itself. The trend is toward adaptive: Claude Opus 5.5, for example, no longer lets developers switch thinking off and instead defaults to medium effort with automatic scaling. For users the practical rule is simple — turn effort up for maths, code, and multi-step analysis, and down for drafting, summarising, and chat.