"What Is Temperature in AI? Controlling LLM Creativity"
Temperature is the single most important parameter for controlling LLM output. Here is what it does and how to set it.
What temperature does
When an LLM generates text, it produces a probability distribution over the vocabulary for the next token. Temperature scales these probabilities before sampling:
- Temperature = 0: Always pick the most likely token (deterministic, “greedy”)
- Temperature = 0.7: Mostly pick likely tokens, but occasionally pick less likely ones (balanced)
- Temperature = 1.0: Sample proportionally to the model’s raw probabilities (default for most models)
- Temperature = 2.0: Nearly random — output is chaotic and often nonsensical
The effect on output
| Temperature | Behavior | Best for |
|---|---|---|
| 0.0 | Deterministic, factual, repetitive | Code, data extraction, factual Q&A |
| 0.3 | Mostly factual, slight variation | Summarization, structured output |
| 0.7 | Balanced — creative but coherent | Chat, general writing, brainstorming |
| 1.0 | More creative, more varied | Creative writing, ideation |
| 1.5+ | Wild, unpredictable, often incoherent | Not recommended |
Practical examples
Temperature 0.0 — code generation: Prompt: “Write a Python function to reverse a string.” Result: Consistent, correct code. Every run produces the same output. Good for reliability.
Temperature 0.7 — blog writing: Prompt: “Write a blog intro about morning routines.” Result: Coherent, slightly varied each run. Good for drafts you will edit.
Temperature 1.0 — creative writing: Prompt: “Write a poem about the ocean.” Result: More varied imagery and word choice. Good for creative tasks.
When to use what
- Code generation / data extraction / factual Q&A: 0.0 (or 0.1)
- Summarization / translation: 0.3
- Chat / general writing / email drafting: 0.7 (default for most APIs)
- Creative writing / brainstorming / ideation: 0.9–1.0
- Never use: 1.5+ (output degrades rapidly)
Related parameters
- Top-p (nucleus sampling): Only sample from the top-p% of probability mass. Top-p = 0.9 means the model considers only the top 90% of likely tokens. Often used with temperature.
- Top-k: Only sample from the top k most likely tokens. Less common in 2026.
- Frequency penalty: Reduce probability of tokens that have already appeared. Prevents repetition.
- Presence penalty: Reduce probability of tokens that have appeared at all. Encourages diversity.
FAQ
What is the default temperature for GPT-4o? 1.0 for chat completions. Many developers override to 0.7 for general use or 0.0 for code.
Does temperature 0 always give the same output? Almost. Floating-point math and batch size can cause tiny variations. For true determinism, use seed parameter (if available) + temperature 0.
Can high temperature cause hallucinations? Indirectly — higher temperature means the model is more likely to pick less probable tokens, which can lead to less factual output. But hallucination is more about the model’s training data and capabilities than temperature.
Verdict
Use temperature 0.0 for code and facts, 0.7 for general writing, and 1.0 for creative tasks. If you are not sure, 0.7 is the safe default. The difference between 0.7 and 1.0 is subtle; the difference between 0.0 and 1.0 is dramatic.