Order prompts static-first/dynamic-last so prompt caching actually fires — 70–90% input cost cut

TokenMix·high signal

Prompt caching only hits when the prefix is byte-identical, so structure every request as system prompt → tool defs → few-shot → document context (all static, cacheable) and put the user query and any changing data dead last. On Anthropic, cached reads cost 10% of normal input (writes cost 1.25×), yielding ~70–90% savings in real workloads — but a single dynamic token near the top silently busts the whole cache. The 2026 cost-optimized stack layers prefix caching + conversation compaction + semantic/plan caching on top.