Frequency Penalty

Frequency penalty discourages the model from repeating tokens in proportion to how often they’ve already appeared in the output. Each time a token is used, its score for future selection drops a bit more — a word used five times is penalized five times harder than a word used once. The frequency_penalty parameter typically ranges from -2.0 to 2.0, with 0 as the default; positive values reduce repetition, and the scaling with count is what distinguishes it from its sibling, presence penalty, which applies one flat penalty regardless of how many repeats occurred.

You reach for this when output gets stuck in loops: the same phrase recycled across paragraphs, a summary that restates its opening line, list items that all begin identically, or — the pathological case, common at low temperature — the model emitting the same token endlessly. Long generations are where repetition compounds, so this knob matters most for articles, product descriptions at scale, and other extended prose. It’s a lighter, cheaper fix than raising temperature, which changes randomness everywhere rather than just discouraging reuse.

In practice, frequency_penalty is a top-level request field on OpenAI-style APIs (Anthropic doesn’t expose one — you steer repetition through prompting instead). Start small, around 0.1 to 0.5, and inspect output before going higher: strong penalties punish necessary repetition too, mangling code, JSON keys, and technical terms that legitimately recur. Keep it at 0 for structured output.