“The primary control lever for modern reasoning models has shifted from 'temperature' to the 'reasoning budget' (the number of tokens allocated for thinking), as providers like OpenAI and Anthropic now force temperature to 1.”