Temperature controls how random a model's outputs are. Lower values stay close to the most likely tokens (good for extraction, classification, and anything that must be reproducible); higher values add variety (good for brainstorming). Prompts that must be consistent should keep temperature low.
Why it matters
Temperature decides whether a prompt is reproducible. High temperature on a task that needs exact output is a leading cause of flaky evaluations and inconsistent production behavior.
How it works
Lower temperature narrows sampling toward the most likely tokens, approaching deterministic output; higher temperature widens it. Set it per use case: low for extraction and classification, higher for brainstorming.