The short answer
Temperature is a setting that controls how random an AI model's word choices are. Low values make it pick the most likely next words, giving focused, repeatable answers; high values allow less likely words, giving more varied output. OpenAI's Responses API lists it from 0 to 2 with a default of 1, but GPT-6 models reject it whenever they reason. Anthropic removed it on Claude Opus 5.5, and Google says Gemini 4 ignores custom values.
| What it controls | Randomness in choosing each next token |
|---|---|
| Low temperature | More focused and repeatable |
| High temperature | More varied, more surprising, more mistakes |
| OpenAI Responses API | 0 to 2, default 1; GPT-6 models accept it only with reasoning effort none, which GPT-6 Astra does not offer |
| Claude Opus 5.5 | Removed; Anthropic says to guide behaviour with prompting |
| Gemini 4 | Custom values ignored, with a deprecation warning |
| In chat apps | Not a setting you can change in ChatGPT, Claude or Gemini |
What temperature means, in plain English
A language model writes by choosing one token (a word or part of a word) at a time. At each step it has a list of candidates with probabilities. Temperature changes how that list is used. At a low temperature, the model almost always picks the top candidate, so the same prompt tends to give the same answer. At a higher temperature, lower-ranked candidates get more of a chance, so answers vary more and can surprise you, for better or worse.
Google's Gemini API documentation puts it this way: temperature controls the degree of randomness in token selection; lower values suit prompts that need a more deterministic, less open-ended response, while higher values can lead to more diverse or creative results. OpenAI's API reference says the same in numbers: values like 0.8 make output more random, values like 0.2 make it more focused and deterministic, and it recommends changing either temperature or a related setting called top_p, but not both.
Where you can (and cannot) set temperature in November 2026
| Where | Temperature setting |
|---|---|
| ChatGPT, Claude and Gemini apps | Not exposed. The app chooses its own settings. |
| OpenAI Responses API, GPT-6 models | Listed as 0 to 2 with a default of 1, but only accepted when reasoning effort is none. GPT-6 Astra has no none setting, so it never accepts temperature. |
| Claude API, Claude Opus 5.5 | Removed, along with top_p and top_k. Anthropic says to use prompting instead. |
| Gemini API, Gemini 4 and Gemini 3.6 or later | Custom values ignored and a deprecation warning returned |
| Gemini API, earlier Gemini 3.x models | Adjustable, but Google strongly recommends the defaults |
The trend is clear. Google warns that on Gemini 3.x models, setting temperature below 1.0 can cause unexpected behaviour such as looping or degraded performance, especially in complex maths or reasoning. Anthropic's migration guide for Claude Opus 5.5 says plainly: omit temperature, top_p and top_k, and use prompting to guide the model's behaviour. OpenAI's guide to the GPT-6 family says to remove temperature and top_p whenever reasoning effort is not none, and GPT-6 Astra and GPT-6.1 Sol do not offer none at all. As reasoning models became the default, the labs moved control from randomness to effort: how much the model thinks before answering.
Why it matters at work
For most people, the honest answer is that temperature is no longer the setting to worry about. You cannot change it in the chat apps, and on the newest models developers cannot either. What still matters is knowing what it was for, because the goals have not gone away:
- When you want consistency, for example classifying support tickets or extracting fields from invoices, the modern tools are clear instructions, a fixed output format and examples in the system prompt, plus structured outputs where the API offers them.
- When you want variety, for example 20 campaign names or five different openings for a talk, ask for it directly: "give me 10 options that are deliberately different from each other".
- When you want accuracy, temperature never guaranteed it. Use a higher effort setting, give the model your source material and check the output. Our guide on reducing AI hallucinations covers what works.
If your team maintains older integrations, check them. Requests to Claude Opus 5.5 should no longer send temperature, OpenAI's GPT-6 models reject it while reasoning, and Gemini 4 ignores it, so code that relied on it may behave differently after a model upgrade. Our guide on choosing the right AI model helps with upgrade decisions.
An example you can try without any settings
You can see the effect of randomness in any chat app, even without the dial:
- Ask ChatGPT, Claude or Gemini the first prompt below in three separate new chats.
- Compare the three answers. The wording will differ even though the prompt is identical. That variation is sampling at work.
- Now use the second prompt, which fixes the format and gives an example. Run it three times and compare again. The answers should be far more alike.
You have just done with instructions what developers used to do with temperature, which is how the labs now recommend you do it.
Common misconceptions about temperature
- "Temperature 0 gives identical answers every time." Google's documentation describes temperature 0 as deterministic, always choosing the highest-probability response, but Anthropic notes that temperature 0 never guaranteed identical outputs on its earlier models. Treat it as "much more consistent", not "guaranteed identical".
- "Low temperature means more accurate." It means more predictable. A model can be predictably wrong. Accuracy comes from good inputs, reasoning and checking.
- "High temperature means more creative." It means more random. Past a point, randomness produces nonsense, not ideas. Asking for varied options usually works better.
- "I can turn the temperature down in ChatGPT." Not in the app. Typing "set temperature to 0" in a chat does not change the setting; at most the model treats it as a request to be more careful.
Related terms
- Token: the unit temperature chooses between.
- Reasoning model: why effort has replaced temperature as the main dial.
- System prompt: the better way to get consistent output.
- Hallucination: the problem temperature does not solve.
- AI API: where developers set temperature on models that still accept it.
- Inference: the process temperature affects.
To put this into practice, see how to write a good prompt, our brainstorming prompts and the guides to ChatGPT, Claude and Gemini.
Prompts to copy
See the variation for yourself
Write a two-sentence description of our new [product or service] for [audience].
You get: Three different answers from three chats, showing sampling at work
Get consistent output without temperature
Write a description of our new [product or service] for [audience]. Format: exactly two sentences. Sentence one states what it is and who it is for. Sentence two gives the single most useful benefit with a number if available. No adjectives such as innovative or powerful. Example of the format: 'Acme Notes is a meeting note-taker for small sales teams. It cuts follow-up writing to under five minutes a call.'
You get: Near-identical structure every time you run it
Get deliberate variety without temperature
Give me 10 names for [project or campaign]. Make them deliberately different from each other: 2 descriptive, 2 playful, 2 one-word, 2 based on a metaphor and 2 that would work in [other language]. For each, add one line on the impression it gives.
You get: A spread of options that covers the range on purpose