What does the dial do to the next word?
At every step, a model gives each possible next word a raw score. Those scores are turned into odds, and one word is drawn according to the odds. Temperature divides the scores before that happens. A low temperature stretches the gaps, so the favourite becomes a near certainty. A high temperature squashes them, so the outsiders get a real chance.
The developer settings say how far the dial goes. OpenAI's API accepts a temperature from 0 to 2 and Anthropic's from 0 to 1, according to their API references, and some newer models do not let you change it at all.
Which jobs want it low, and which want it high?
Low suits anything with one right answer: pulling figures out of a report, sorting enquiries into categories, a reply that quotes your refund policy. Higher suits the jobs where you want options to choose from, like names for a new product, ten headline angles for an ad or a fresh way into a tired newsletter. A middle setting is the usual default for general chat, because it reads naturally without wandering.
What happens to a winery's tasting offer at two settings? (example figures)
A cellar door near Mudgee charges $15 for a tasting flight of six wines, waived when you buy two bottles. The owner asks a model, through a tool with a temperature setting, for five versions of an Instagram caption, once at 0.2 and once at 1.2.
| Temperature 0.2 | Temperature 1.2 | |
|---|---|---|
| How different the five are | Barely: four open with the same line | Five angles: the view, the shiraz, the drive, the cheese board, the long lunch |
| Price | $15 in all five | $15 in four, missing in one |
| The offer | Waived with two bottles, all five | One says "free with any bottle" |
| Usable | One, because the rest repeat it | Three, after fixing the offer |
The creative set is better, and one caption in it gives the tasting away for one bottle instead of two. If that version runs and 40 visitors over a long weekend buy a single bottle and ask for their $15 back, the cellar door either refunds $600 it never offered or has 40 awkward conversations. Facts belong in the prompt, and every version gets checked against them whatever the setting.
Can you change the temperature in ChatGPT or Gemini?
Not in the everyday chat apps, which set it for you. Developer tools such as Google AI Studio and the OpenAI Playground show it as a slider. In an ordinary chat you can get the same effect through the prompt: "give me ten very different options" pushes for variety, and "use exactly these figures and do not change them" pushes the other way.
Which part of Cora's advice must never vary?
The figures: for anyone working with an AI like CoraVision, temperature is a reminder that the wording of advice can vary while the figures must not, and CoraVision takes the figures from the accounts themselves. Once Google Ads, Meta and Google Analytics 4 are connected, it reads them the day they are connected and nightly from then on. The one number your 12-month plan is built around stays put while the sentences around it change, and the weekly check-ins measure progress against that number.
Questions
Does a temperature of 0 make an AI accurate?
No, it makes it consistent. A model at 0 picks its most likely word every time, so it can give the same wrong answer every time. Accuracy comes from the facts in the prompt and the checking afterwards.
Why do I get a different answer when I ask the same thing twice?
Chat apps run above zero, so each run picks its words afresh and takes a slightly different path. The app may also have been updated between your questions. That is handy for ideas, and for facts a changing answer is a sign to check.