Both major labs converged on the same control by 2026: Anthropic exposes effort at low, medium, high, xhigh, and max, and OpenAI exposes a reasoning-effort setting running from none to max. It is a behavioural signal rather than a hard token budget, and on Claude models it affects all output tokens, including how many tool calls the model makes, not just its thinking.
Its arrival changed what belongs in a prompt. Instructions like “think step by step” or “be thorough” were prose attempts at a dial that now exists, and carrying effort defaults across a model generation is unreliable because a given level buys more capability on each new release. The usual advice is to re-run an effort sweep on your own evaluations after upgrading rather than reusing the previous setting.