Temperature

At each step, a model assigns a probability to every possible next token. Temperature rescales those probabilities before one is sampled: values below 1 concentrate the distribution on the top choices, and values above 1 flatten it. At or near zero, output becomes close to deterministic. It is often used alongside other sampling settings such as top-p (nucleus sampling), described by Holtzman et al. (2019).

Why it matters: the right temperature depends on the task. Data extraction, classification and code usually benefit from a low temperature for repeatable results; brainstorming and creative writing can use higher values. Products often set it for you, which partly explains why the same model feels different across tools.

Example: asking for ten product name ideas at a low temperature may return similar, safe names; a higher temperature produces more unusual options.

Limits: a low temperature makes answers more consistent, not more accurate. Some reasoning models fix or restrict temperature, and the valid range differs between providers.