Presence Penalty

Presence penalty applies a one-time, flat penalty to any token that has appeared in the output at all — used once or used fifty times, the penalty is identical. That’s the precise contrast with frequency penalty, which stacks with every repetition. The effect is different in kind: frequency penalty says “stop overusing that word,” while presence_penalty says “you’ve covered that — move on to something new.” It ranges from -2.0 to 2.0 on OpenAI-style APIs, defaulting to 0, with positive values pushing the model toward fresh vocabulary and, by extension, fresh topics.

This is the knob for breadth. When you’re brainstorming and every idea circles back to the same theme, when a chatbot keeps steering conversations to one subject, or when generated variations all feel like the same answer reworded, a positive presence penalty nudges the model into unexplored territory. It’s less about polish (frequency penalty’s job) and more about diversity of content — which also means it’s the wrong tool when consistency is the goal: penalizing a key term’s first reuse can push a focused answer off-topic.

In practice, presence_penalty is a top-level request field alongside temperature and frequency_penalty (again, Anthropic omits it — prompt for variety instead, e.g. “give five ideas from genuinely different angles”). Try 0.3 to 0.8 for ideation and creative variety, keep it at 0 for extraction, code, and structured output, and tune one repetition knob at a time so you can tell which one changed the behavior.