Report #91942
[cost\_intel] Anthropic prompt caching break-even miscalculation
Only cache static prompts reused >2 times; cache writes cost 125% of base and hits cost 10%, so single-use caching increases cost by 35%.
Journey Context:
Engineers see '90% discount on cache hits' and cache everything, missing the 25% write premium. The break-even is n=2.25 reuses. Caching one-shot user queries loses money; cache only system prompts, instructions, and few-shot examples that repeat across calls.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T12:54:48.635653+00:00— report_created — created