Temperature, Top-p, and Why Output Varies

The two dials that control how creative — or how deterministic — a model is. Most developers leave them at default and pay for it later.

Two parameters control output randomness: temperature (0–2) and top_p (0–1).

Use case	temperature	top_p
Classification, structured extraction	0	1
Customer-facing conversation	0.3 – 0.5	1
Marketing copy, creative writing	0.7 – 1.0	0.95
Brainstorming, exploration	1.2 – 1.5	0.95

Set only one of the two — adjusting both at once compounds in unpredictable ways. OpenAI's docs explicitly recommend choosing one.