Temperature

Temperature is a number (typically 0 to 2) that scales the model's probability distribution before sampling the next token. At temperature 0, the model always picks the highest-probability token — output is repeatable and conservative. At temperature 1, you sample from the model's natural distribution — varied but coherent. Above 1, you flatten the distribution toward uniform, which produces increasingly weird and ungrounded text. In practice, the useful range is 0 to about 1.2. Use 0 (or near-0) for code generation, structured extraction, classification, factual Q&A, and anything where you want the same input to produce the same output. Use 0.7-1.0 for creative writing, brainstorming, varied responses, and chat. Use higher only when you specifically want chaos — generating diverse synthetic data, breaking out of repetitive loops. Temperature 0 is not actually deterministic in most production APIs: floating-point non-determinism, batch effects, and routing across hardware mean the same prompt at temp 0 can still vary slightly across calls. If you need bit-exact reproducibility, you typically need a seed parameter (where supported) or a self-hosted model. Some 2026 reasoning models (GPT-5 Thinking, Claude with extended thinking) ignore temperature for the reasoning trace itself and only apply it to the final output.