A reasoning budget is the cap on the tokens a model may use for internal thinking before it produces an answer. It does not decide how long the answer is, only how long the model may think. Raising the budget will not automatically make the answer longer; setting it too low may mean a hard question gets answered before it has been thought through.
The three vendors do not use the same names, so do not mix them up
OpenAI uses reasoning.effort, with tiers of none, minimal, low, medium and high; a tier is not a token count. Claude's older style uses budget_tokens inside thinking, a number you enter directly, with a minimum of 1,024 and a value that must be below max_tokens; the newer style uses output_config.effort, with tiers of low, medium, high, xhigh and max, and the old form has been deprecated on newer models. Gemini 2.5 uses thinkingBudget, a number where 0 switches thinking off and -1 means dynamic thinking; Gemini 3 uses the thinking_level tier of minimal, low, medium or high. Ordinary users in a chat app can usually only pick a model, switch thinking on or off, or choose a tier; entering an exact number is a job for developers calling the API.
Raising or lowering it, and what each costs
Thinking tokens are billed as output tokens, and they also use up context and the output limit. Set the budget too high and too little room is left for the answer, so the model may be cut off before it finishes writing; set it too low and hard questions may skip steps or go wrong. Raising it usually helps with maths, multi-step planning and complex debugging, but costs more and takes longer; for looking things up, summarising, classifying or formatting, the gain is small. Lowering it suits live conversation, voice and bulk processing: cheaper and faster.
Three common misunderstandings
First, a budget is a cap, not an exact amount. Setting 8,000 does not mean 8,000 is spent every time; a simple question makes the model stop thinking early. Second, raising it does not always mean more accuracy; returns diminish, and overthinking can go down the wrong path or talk itself out of a right answer. Third, the full thinking process is not visible by default; you only see a summary or an expandable thinking panel. Start from the lowest tier for each task and raise it step by step until you find the tier that is enough without waste; that is the right way to use a reasoning budget.
