What reordering a prompt so the static prefix goes first saves at cache-read rates, without deleting a token.
Today, an unstable prefix pays full input price on every request. Refactored, the first request writes the static prefix at the cache-write rate and every later request reads it at the cache-read rate, paying full price only on the dynamic tail.
Yes. The prefix cache is byte-exact from token 1. A single changed token at position 100 invalidates everything after it, so a timestamp near the top of the system prompt silently uncaches the whole prompt, including tool schemas and reference docs below it.
Only the stable prefix caches. If dynamic tokens outweigh static ones, most of every request still bills at full input price. Summarize volatile content or externalize state to a tool call so it never enters the prompt.
Slightly. The first request pays the cache-write rate on the static prefix instead of the input rate - a one-time premium amortized to nothing across thousands of requests.
Trimming tokens saves once per request; ordering saves on every request without removing anything. You keep the full reference docs, and the static prefix still bills - at the cache-read rate instead of the input rate.
I build and ship the ClaudFlow workflow builder, including its eight block types and JSON export format. The agent loop patterns these calculators price are patterns implemented in shipped code, not invented for a blog post. Everything runs client side by design: nothing you type leaves this page. More on the about page.
ClaudFlow is a free, privacy-first suite of AI workflow tools. No accounts, no tracking, no data leaves your browser. Part of the Zovo Tools network by Michael Lip. Token counts are estimates, and actual usage depends on encoding, model version, and prompt structure.