Temperature

What Temperature Controls

When a language model generates the next word in a response, it assigns probabilities to thousands of possible options. Temperature determines how it picks from that distribution.

• Low temperature (near 0) — the model almost always picks the highest-probability option. Outputs are consistent, predictable, and focused.
• High temperature (1 or above) — the model samples more freely from lower-probability options too. Outputs become more varied, surprising, and sometimes unexpected.

The Scale in Practice

Most AI APIs accept a temperature between 0 and 2. The sweet spots:

• 0–0.3 — use for factual Q&A, data extraction, coding tasks, anything where you want the most likely correct answer consistently
• 0.5–0.8 — a good default for most writing and conversational tasks; creative but still coherent
• 1.0+ — use for brainstorming, creative writing, generating diverse ideas; accepts that some outputs will be off

At temperature 0, the model is deterministic — give it the same prompt twice and you'll get the same response. At high temperatures, the same prompt produces noticeably different results each run.

A Practical Example

Ask a model "What is the capital of France?" at temperature 0 and temperature 1. At 0: "Paris" every time. At 1: probably still "Paris," but occasionally it might go on tangents or add unexpected framing — because the model is sampling more loosely.

For most AI app development, starting with a temperature of 0.7 and adjusting based on whether outputs feel too repetitive or too unpredictable is a reasonable approach.