Agentic AI Glossary

Cache tokens

Input tokens a provider can reuse from an earlier request instead of processing again.

Illustrative conversation

  1. Why does the usage report show 8,000 input tokens but only 6,000 cached tokens?
  2. The shared prefix matched the cache; the remaining input was new or changed and still had to be processed normally.

Definition

Cached tokens are input tokens whose previously computed prompt state a provider can reuse when a later request has a matching prefix. They are still part of the request and context window, but a cache hit can reduce input-processing cost and latency. The model generates a fresh output, so caching does not make identical prompts return identical answers.

Reviewed