Temperature

Temperature scales the probability distribution over next tokens before sampling. At temperature 0, the model always picks the highest-probability token (deterministic, boring, consistent). At 1.0, sampling matches the model's learned distribution (balanced). At 2.0, distribution flattens, output becomes erratic. Most production code uses temperature 0.0 to 0.7.

Math: logits are divided by temperature before softmax. Lower temperature sharpens the distribution; higher flattens it. Temperature 0 is not technically in the formula (would divide by zero) · it's implemented as argmax (pick highest logit). Common uses: 0 for structured extraction, factual Q&A; 0.3-0.7 for balanced chat; 0.8-1.2 for creative writing. Temperature and top-p are often used together but shouldn't both be tuned aggressively · pick one.

Temperature = 0 breaks ties by tokenization order in most implementations, which is why some "deterministic" calls still vary across infrastructure. True determinism requires seed parameter (OpenAI supports, Anthropic doesn't). Temperature affects reasoning quality non-linearly · too low produces brittle reasoning, too high introduces errors. For reasoning models, temperature usually sets at 0-0.3 range for correctness.