Prompt Compression: Say the Same Thing in Fewer Tokens
Much of a typical prompt is filler the model does not need. Prompt compression strips that redundancy so you send fewer input tokens for the same result — a direct cut to the expensive input side.
Where the waste hides
Verbose instructions, duplicated context, boilerplate and over-long examples inflate input with little benefit. Tightening wording and removing repetition often removes a meaningful share of input tokens on every call.
Compress context, then summarize history
For retrieved context, keep only the passages that matter. For conversations, replace old turns with a running summary. Both shrink input without the model noticing the difference in most tasks.
Calculators for this topic
Frequently asked questions
Does prompt compression hurt quality?+
Done well it removes redundancy the model does not need, so quality holds while input tokens drop.
How much can compression save?+
It varies with how verbose your prompts are, but trimming filler and duplicated context often cuts input noticeably.