Automation and AI

Temperature

Also called Sampling temperature

A setting that controls how predictable a model's wording is, from tightly repeatable output to loose and varied.

Quick facts: Temperature

Category
Automation and AI
Also called
Sampling temperature
Level
Advanced
Affects
Output variety, repeatability, format reliability
Where to see it
Model APIs, provider playgrounds, automation platform nodes
In this article4
  1. What temperature controls
  2. Why temperature matters
  3. Common mistakes with temperature
  4. How to act on it

What temperature controls

At every step of writing, a model has a ranked set of possible next pieces of text, some far more likely than others. Temperature decides how strictly it sticks to the front of that ranking. Turned low, it takes the most probable option almost every time, so the same request produces close to the same answer twice over. Turned high, less likely options get a real chance, and the output becomes more varied and less predictable.

Two things are worth being clear about. First, it changes the wording, not the knowledge — the model does not know more at one setting than another. Second, most chat interfaces do not expose it at all; you usually meet it in an API, a playground panel, or a node in an automation platform. A related control, top-p, restricts the pool of candidates rather than reshaping the odds, and providers generally suggest adjusting one or the other rather than both.

Why temperature matters

It decides whether a task is repeatable. Anything that feeds into a process — sorting enquiries into categories, extracting fields from an invoice, generating a meta description in a fixed format, returning structured data another system will read — needs the same input to give the same output. A low setting is what makes that possible, and it is what lets you test a change and trust the comparison.

At the other end, variety is the point. Generating a batch of ad headline ideas, brainstorming campaign angles or drafting alternative subject lines all suffer at a low setting, where the model returns near-identical phrasings and you end up testing the same idea several times over.

Common mistakes with temperature

Treating it as an accuracy control is the main one. Lowering it makes output more consistent, not more truthful: a model will state the same invented detail confidently every time, and consistency can make a wrong answer look more credible. If facts are the concern, supply the source material and check the output — the setting will not save you.

Running everything high is the second, usually because a demonstration at a creative setting looked impressive. It produces drift: instructions loosely followed, formats broken, the occasional sentence that wanders off the brief. The third is changing it while also changing the prompt, which leaves you unable to tell which change caused what.

How to act on it

Match the setting to the job. Keep it low for anything structured, repeated, or destined for another system, and raise it only where you genuinely want options to choose between. If output that must be reliable is coming back inconsistent, look at the setting before rewriting the prompt.

Change one thing at a time. Fix the wording, test, then adjust the setting and test again on the same inputs, so you can attribute the difference. Write down the setting alongside the prompt in whatever template you save, because a workflow that behaves differently on a colleague’s machine is often just a different value here. In any automated workflow, treat it as configuration to be recorded rather than a knob to fiddle with.

Do and do not

Do

  • Keep it low for classification, extraction and structured output
  • Raise it only when you want options to choose between
  • Record the setting alongside the prompt in your template

Do not

  • Expect a low setting to make answers more truthful
  • Change the setting and the prompt at the same time
  • Leave a platform default in place for repeatable tasks

Questions people ask about this

What temperature should I use?

It depends entirely on the job. Use a low setting where the same input must produce the same output — classification, data extraction, structured formats, anything another system will read. Use a higher one where you want several different options to choose between, such as headline or subject line ideas. Test on your own task rather than copying a value from a tutorial.

Will lowering the temperature stop the model making things up?

No. It makes the output more predictable, not more accurate. A model at a low setting will repeat the same invented detail every time, which can make an error look more trustworthy rather than less. Reducing invention is a matter of supplying the facts with the request, asking it to say when it does not know, and checking what comes back.

Why is my automation giving different answers to the same input?

A higher temperature setting is the usual cause, and it is often left at a platform default. Check the value in the node or API call before assuming the prompt is at fault. Model updates and other sampling settings can also contribute, so record the configuration alongside the prompt and re-test with a fixed set of example inputs after any change.

Related terms

Found this useful?

Share it, or ask an AI to summarise it

Back to the glossary

Knowing the term is the easy part

Applying it to your own site and budget is the work. Book a call and I will tell you what actually applies to you.