Prompt Cache Explorer

LLM providers can reuse the work already done for a prompt’s beginning. Type prompts, press Check, and see which parts came from the cache.

1 · Write a prompt

2 · Result

Cached — read from cache, not recomputed Not cached — computed now, then stored
Your prompt will appear here, colored by cache status.

*Illustrative: cached tokens billed at 10% of the normal input price. Real prices vary by provider.

The cache

Each box is one block of tokens, stored together with everything before it. Boxes nested under another extend that prefix.

Empty — check a prompt to fill it.

How it works

The model processes a prompt token by token, building internal state (the KV cache) for each one. If a new prompt starts with exactly the same tokens as an earlier one, that state can be reused instead of recomputed.

Two consequences you can test: (1) only a shared prefix counts — an identical sentence in the middle of a prompt gets no cache hit if anything before it differs, and (2) caching works in blocks (try the slider): a partial final block is stored, but only an identical prompt can reuse it. At block size 1 it is a pure token-by-token prefix match.

Practical tip: put stable content (system prompt, documents, examples) first and the changing question last.