Prompt Caching: Cut Latency and Cost by Reusing Context

Prompt caching lets the model reuse the unchanging front of your prompt — system instructions, tool schemas, a long reference document, or a storyboard style guide — instead of re-reading it on every call. On a cache hit you pay a fraction of the input price and get your first token noticeably faster, which matters most for long system prompts and multi-turn chats where the same 10k-token preamble repeats.

To use it well, put stable content first and volatile content last: the cache breaks at the point the tokens diverge, so a single edited character near the top invalidates everything after it. Mark the boundary right after your largest reusable block, keep it above the recommended minimum token length, and remember caches expire after a few minutes of inactivity — batch related requests to stay warm.

On CinderHub, caching a shared character sheet or house art-direction prompt across a whole image or video batch keeps every frame on-style while slashing per-render cost. Measure your cache hit rate, not just the config: if it stays low, your prompt is probably shuffling supposedly-static content and quietly forcing full re-reads.