Dictionary term
Temperature
Temperature is a sampling setting that controls randomness in a model's output. The allowed range depends on the provider (for example 0 to 2 on OpenAI's Responses API). Lower values favour the most likely tokens and higher values allow more varied choices, but even a temperature of 0 does not make API output fully deterministic.
In plain English
Temperature is a setting that makes an AI's answers more predictable (low) or more varied (high). Many of the newest models no longer let you change it.
Analogy
Like a dial between a careful accountant (low) and a freewheeling brainstormer (high).
Why it matters
On models that still accept it (many OpenAI models, older models, and self-hosted open-weight models), a low temperature keeps extraction and classification more consistent. Vendors are removing it from their newest models: Anthropic's API version 2026-10-21 returns a 400 error if temperature is sent at all, Google says its newest Gemini models ignore custom values, and OpenAI's current flagship does not accept a custom temperature. On those models, get consistency from clear instructions and structured outputs instead.
In practice
When extracting a phone number with an open-weight model served on vLLM, set temperature near 0. When the same step runs on a current Claude model, drop the parameter and enforce the format with a JSON schema.
Related terms
Sources
Browse all terms →Last verified