Temperature

What is a Temperature

Temperature is a setting that controls how much randomness goes into a model choosing its next token. At low temperature, the model strongly favors whatever token it considers most likely, producing output that is more consistent and predictable. At higher temperature, less likely tokens get picked more often, producing output with more variety, and sometimes more creativity, at the cost of consistency.

This is a request-level setting, not a fixed property of the model. A task like generating code or extracting structured data usually wants low temperature, since consistency and correctness matter more than variety. A task like brainstorming options or generating creative copy can benefit from higher temperature, where some randomness surfaces ideas a fully deterministic pass would not.

Temperature is not a fix for hallucination, and low temperature does not make a model's factual claims more accurate, it only makes its phrasing and token choices more consistent. A model at temperature zero can still confidently state something false, it will just state the same false thing the same way each time, rather than varying it.