Pilot: Prompt caching › Ep 3
Ep 3 · The prefix
Length 1:16Today's goal: The cache only works on the start of your request: the prefix.
What you'll learn
- Send the same conversation plus one new message, and the prefix is read from the cache. Only the new part needs attention computed.
- Reading from the cache is roughly 90% cheaper. Not free, just cheap, and writing to the cache costs a little too.
- Change one block near the top, and everything after it is paid again.
Sources
- Burke Holland, “Your AI credits are vanishing - here's the fix” (YouTube): the prefix, and the ~90% read discount
- Anthropic's prompt-caching reference (as bundled with Claude Code 2.1.299, 21 Oct 2026): prefix match and prices