Turning temperature up doesn’t make the model smarter, and turning it down doesn’t make it more truthful. It only changes how much variety the model allows when it picks each next word. Lower means more expected, repeatable wording. Higher gives it more room to vary.
Think about asking someone to answer from a script versus asking them to brainstorm. Confirming office hours, you want the script. Naming a workshop, you want more range. Same person, different amount of freedom.
How it shows up
For work that needs consistency (classifying transactions, extracting invoice fields, following a strict format) you usually want low temperature, the same kind of answer every time. For naming ideas, first drafts, or different ways to explain a concept, a little more helps you explore, and you still review the work. Most clients won’t touch it directly; it shows up in APIs, custom agents, and advanced settings, and it’s one reason the same prompt won’t give the identical output every time.
Why you care
More variation isn’t real creativity. The model is still predicting language from its training and your prompt, so temperature changes how it picks words, not whether it has taste or judgment. It also doesn’t fix hallucination, since a low-temperature answer can still be wrong without the right context. Use low freedom when work needs precision, more when it needs options, then apply reasoning and human review either way.