Temperature affects the randomness of token selection. Lower values generally make a model’s responses more predictable; higher values generally allow more varied continuations. The practical effect depends on the model, sampling implementation and task, so a temperature value does not guarantee a particular level of creativity or accuracy.
A support classifier that must return the same label for the same ticket usually benefits from a low temperature, while a tool that suggests ten blog title ideas can use a higher one. Some models restrict or ignore this setting, and some newer models expose different controls, so check the provider documentation. For format-critical work, pair a low setting with structured output.
For extraction, classification or tightly formatted output, test a low setting and measure errors. For brainstorming, variation may be useful. Keep the setting consistent during evaluations so results can be compared fairly. Related: Top-p Sampling, LLM Evaluation.