AI glossary
Temperature
A setting that controls how random a model’s word choices are: low values give predictable text, high values more varied text.
At each step the model has a probability for every possible next token. Temperature reshapes those probabilities. Near 0, it almost always picks the most likely token, which suits extraction or factual tasks. Higher values let less likely tokens through, which can help brainstorming but increases mistakes.
Some recent models fix sampling settings and offer other controls instead, such as how much effort they spend reasoning.
Example: At low temperature, “suggest a name for a bakery” gives almost the same name every time. At high temperature it suggests a different one each time: some more original, some nonsensical.
In practice
- For extracting data, classifying or answering factual questions, use low values.
- For brainstorming, raise it a little and ask for several options.
- On Claude models from Opus 4.7 onwards, the API no longer accepts a changed temperature: you steer the result with the prompt and the effort level.