Temperature

At each step a model assigns probabilities to possible next tokens. Temperature adjusts how it picks among them. A low temperature favors the most likely tokens, giving consistent and focused output. A higher temperature spreads the choice more widely, giving more variety.

Lower settings suit extraction, classification, and code, where you want the same input to give the same kind of answer. Higher settings can help with brainstorming and creative writing.

Temperature is a minor lever compared with clear instructions and good context. Even at its lowest, output is not guaranteed to be identical across runs, and some models and modes, particularly those using extended reasoning, restrict which sampling settings you can change.