AI Glossary Term
What does “Frequency Penalty” mean?
Definition
A decoding setting that reduces the probability of tokens proportionally to how often they've appeared so far.
How Frequency Penalty works
Example
If a model has already written 'amazing' four times, a frequency penalty makes a fifth use less likely than the first was, nudging it toward other words.
Why it matters
It reduces repetitive wording in long outputs. High values can push the model toward odd or unnatural word choices.
Common questions
Frequency vs presence penalty?
Frequency scales with the number of uses; presence applies a flat penalty once a word has appeared at all.
Which providers offer it?
Several API providers expose it as a setting, with differing ranges.
Should I always use it?
Usually no. The default is often fine, and it is mainly for cases with noticeable repetition.