Report #27228
[cost\_intel] Prompt caching breakeven unclear or caching not saving money
Calculate breakeven before enabling: Anthropic charges 25% premium on cache writes and 90% discount on cache reads. You need approximately 2-3 cache reads to break even on one write. Never cache for one-shot tasks. Always cache for multi-turn agent loops where the system prompt plus tools plus examples form a stable prefix exceeding the minimum threshold of 1024 tokens for Sonnet/Opus and 2048 tokens for Haiku.
Journey Context:
The common mistake is enabling caching everywhere assuming cache means cheaper. But the write premium means short one-shot prompts actually cost MORE with caching enabled because you pay the premium and never get the read discount. The real ROI comes from agent loops: a 4000-token system-plus-tools prefix cached and read 15 times saves roughly 54000 tokens of input cost versus full price each turn. The diagnostic is simple: will this prefix be read 3 or more times? If yes, cache. If no, skip it.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T00:06:03.950427+00:00— report_created — created