The model never picks a word. It picks from a list
After reading your prompt, a language model produces a score for every token in its vocabulary, 100,000 to 200,000 on current models. It then turns those scores into probabilities: "leaves" 45 percent, "weather" 25, "colours" 15, and so on down a long tail. What is an LLM covers why it works that way. Then it draws one token from that list, appends it, and does it all again.
Temperature is applied between the scores and the draw. Every score is divided by the temperature before the probabilities are computed. Divide by a number below 1 and the gaps between scores widen, so the favourite pulls further ahead. Divide by a number above 1 and the gaps shrink, so the long tail gets a real chance. At exactly 0 the maths collapses to "always take the top score", which is why 0 is sold as deterministic.
The name is borrowed from physics, where a hotter gas has more randomly moving particles. A hotter model has more randomly chosen words. Nothing in the model's knowledge changes; only the draw does.
What the number does to the odds, worked through
Take a prompt whose next word the model rates as five candidates at 45, 25, 15, 10 and 5 percent. That's the distribution at temperature 1. Apply the formula at 0.5 and at 2 and the same five candidates come out as below. The percentages are computed from the formula, not measured, and the shape is what every real distribution does.
Read the fifth candidate across the three panels: under 1 percent, then 5, then 11. Over a 500-token answer, an 11 percent chance per token of taking the fifth-best word compounds into text that wanders, invents and eventually stops making sense, which is what people see when they push a playground slider to 2. And read the top candidate: even at 0.5 it loses one draw in three, so a low temperature is consistent, not fixed.
That last point matters for the word "deterministic". Anthropic's API reference says that even at 0 the results "will not be fully deterministic", and the hedge is real. Floating-point arithmetic on parallel hardware and provider-side batching mean the same prompt at 0 can still produce two different answers. If you need identical output, save it; don't rely on the setting.
What each provider allows, and what the defaults are
| Provider | Range | Default | Notes |
|---|---|---|---|
| OpenAI | 0 to 2 | 1 | Reasoning models (the o-series, GPT-5 with reasoning) do not accept the parameter |
| Anthropic | 0 to 1 | 1 | Rejects requests that set both on Opus 4.1 and later |
| Google Gemini | 0 to 2 | 1 | Exposed as a slider in AI Studio |
| OpenRouter | Passes through | Provider default | The route Whizi uses to reach every model |
Three things stand out. Every provider defaults to 1, which means sample from the model's own odds, so the answers you get in an ordinary chat are already one draw among many. Anthropic's scale stops at 1, so a Claude temperature of 1 is the ceiling while an OpenAI 1 is the midpoint; the numbers aren't comparable across vendors. And reasoning models reject the setting: OpenAI lists temperature as unsupported for the o-series and GPT-5 and returns an error if you send it, and Anthropic requires 1 whenever extended thinking is on.
Which setting for which task
The rule is one line: lower for tasks with a right answer, higher for tasks with many good answers.
| Task | Temperature | Why |
|---|---|---|
| Extraction into a table or JSON, classification, code | 0 to 0.3 | One right shape; variation is a defect |
| Summaries, translation, factual questions | 0.3 to 0.7 | Some phrasing freedom, no invention wanted |
| Drafting, rewriting, brainstorming | 0.7 to 1 | The provider default; varied but coherent |
| Fiction, names, slogans, deliberately odd ideas | 1 to 1.3 | Long-tail words are the point |
| Above 1.5 | Rarely | Coherence drops before novelty rises |
Two cautions. Temperature does not control truth. A model at 0 picks its most probable answer, and if the most probable answer is wrong, it's wrong every time with total confidence, which what is an AI hallucination explains from the training side. And temperature isn't the lever for tone or length. Those belong in the instructions, and prompt engineering for beginners covers them; the number only decides how often the model gambles on a word.
Why you have never seen the slider in a chat app
ChatGPT, the Claude app, Gemini and Whizi expose no temperature control. The setting lives in developer playgrounds and APIs, and none of the three vendors publishes the value its own app uses. Whizi sends none, so every model here runs at its provider default of 1. The reason is the one above: the default is already the setting most tasks want, and a slider in the wrong place turns a good assistant into a bad one with no signal about which knob did it.
What a multi-model workspace gives you instead is a bigger lever. Two different models are two different probability lists, and switching between them changes the answer far more than any temperature would. The habit that replaces the slider: for a task with one right answer, ask a second model and act only where they agree; for a task with many good answers, ask two models and pick. In Whizi that's a model change mid-conversation, and switching models mid-conversation covers what carries across.
If you do need the knob, it's in the provider's own API or playground, per request, and the ranges in the table above apply.
- Use 0 to 0.3 for anything with one right shape: extraction, classification, code
- Leave the default of 1 for drafting and rewriting; it is what every provider tuned for
- Never treat 0 as a guarantee of identical output; save the answer you want to keep
- Do not reach for temperature to fix tone, length or accuracy; those are instruction problems
- When a chat app has no slider, change the model instead, which moves the answer more than any temperature
Frequently asked questions
What does temperature 0 mean in AI?
At temperature 0 the model always takes its single most likely next token instead of sampling, so the same prompt gives the same answer nearly every time. Anthropic's API reference says the results "will not be fully deterministic" even at 0, because hardware and batching can still introduce tiny differences. It's the right setting for extraction, classification and code, and it does nothing to make the most likely answer more correct.
What is the best temperature for ChatGPT?
The ChatGPT app doesn't let you set one, and OpenAI doesn't publish the value the app uses. Through the API, use 0 to 0.3 for structured output and factual extraction, 0.7 to 1 for writing, and above 1 only for deliberately varied creative text. OpenAI's reasoning models return an error if you send the parameter at all.
What is the difference between temperature and top-p?
Temperature reshapes the whole probability list, sharpening or flattening it. Top-p cuts the list off, keeping only the most likely tokens whose probabilities sum to p and drawing from those. Both reduce or increase randomness. OpenAI advises changing one or the other, and Anthropic rejects a request that sets both on Opus 4.1 and later.
Does a higher temperature make AI more creative?
It makes the model pick less likely words more often, which reads as creative up to about 1.3 and as incoherent past about 1.5. In a worked example, raising temperature from 1 to 2 lifts the fifth-best candidate from a 5 percent chance to 11 percent on every token, and that compounds over a long answer. Creativity in the sense of a better idea comes from the prompt and the model, not from the dial.
Can I change the temperature in Whizi?
No. Whizi sends no temperature, so every model runs at its provider default; ChatGPT, the Claude app and Gemini expose no control either and publish no value. The equivalent lever in a multi-model workspace is the model picker: two models are two different probability distributions, and switching mid-conversation changes an answer more than any temperature setting would.
