Parameters
Temperature
Temperature scales the model’s output probabilities before sampling: below one sharpens the distribution towards the likeliest token, above one flattens it.
At or near zero the output is effectively deterministic, which is what you want for classification, extraction and anything your code parses. Around 0.7 gives variety while keeping unlikely tokens suppressed.
Above about 1.5 the distribution approaches uniform and output becomes frequently incoherent.
In practice
The default is not neutral, it is a choice someone else made for a general audience. Anything that must be reproducible — classification, extraction, routing, and any grader — belongs at zero. The value is also not calibrated across model families, so a setting carried over from another provider is a guess wearing a number.
Common questions
What temperature should I use?
Zero for anything that must be reproducible — classification, extraction, routing, and any grader. Higher only where variety is the point, such as drafting alternatives. The default is a general-purpose compromise, not a neutral choice.
Does temperature zero guarantee identical output?
No. It removes sampling randomness but not infrastructure variation, batching differences or a model version changing behind an alias. Treat it as close to deterministic rather than as a contract.