What is temperature? The AI's creativity dial

The direct answer

Temperature is a setting that controls how random an AI model's answers are. Turn it down and the model plays it safe, picking the most likely word every time. Turn it up and it takes more chances, giving you more varied — and sometimes stranger — answers.

If you've ever wondered why asking the same chatbot the same question twice gives slightly different answers, temperature is a big part of the explanation. Most chat apps hide this dial from you, but it's one of the main controls developers get when they build on top of a model.

How it works

An AI model writes by picking one word at a time (technically, one token). At each step it assigns odds to thousands of possible next words. Temperature reshapes those odds.

At temperature 0, the model always takes the top pick — the single most likely word. That's called greedy sampling: predictable, repeatable, boring in a good way. Raise the temperature and the odds flatten out, so runners-up start winning sometimes. Higher still, and genuine long shots can win — which is where the weird and wonderful answers come from.

A simple example

As an illustration — not a real test, just the idea. Imagine asking for a name for a lemonade stand.

At low temperature, you'd likely get something sensible like "Sunny's Lemonade." At high temperature, you might get "The Citrus Rebellion" or "Sour Power Hour" — more surprising, more fun, and occasionally complete nonsense. Same model, same question, different roll of the dice.

That's the tradeoff in a nutshell: low temperature buys reliability, high temperature buys variety.

Why it matters

When you want a dependable answer — a summary of a document, a chunk of code, extracted data — low temperature is your friend. It keeps the model close to the most probable, well-trodden path, which is where correct answers usually live.

When you want ideas — brainstorming names, plot twists, ten different angles on a problem — a higher temperature helps the model wander off the beaten path. Many writing and ideation tools expose this as a "creativity" slider, which is really just temperature with a friendlier label.

One more thing worth knowing: temperature doesn't change what the model knows. A model can't become smarter or learn new facts by turning the dial. It only changes how adventurously it picks among the words it already knows.

The common misunderstanding

The common misunderstanding: higher temperature means more creative, so higher is better for creative work. Not quite. Temperature doesn't add imagination; it adds noise. Crank it too high and you don't get brilliance — you get incoherence, because the model starts picking words with barely any connection to what came before.

Real quality in AI output mostly comes from a good prompt plus a model that understands the task. Temperature just decides how much variety you get across tries. If the first answer is wrong, turning up the temperature and asking again gives you a different wrong answer, not a better one.

What changed recently

Two honest footnotes. First, even temperature 0 isn't perfectly repeatable: Anthropic's own documentation notes that identical inputs can still produce different outputs across API calls, thanks to the realities of large-scale computing.

Second, not every model treats the dial the same way. In one recent long-context evaluation, the researchers noted that a reasoning model (o3) simply didn't support the temperature setting at all, so they ran it at its default. As reasoning-focused models spread, temperature is becoming one control among several — not the universal dial it once was.