Prompt Cache Economics — 為什麼你的 AI 帳單比想像中貴

Prompt caching 本來應該幫你省 90% 的 token 費用,但有一個 bug 可以讓你不知不覺多付十倍錢。從 Claude Code 原始碼洩漏的 DANGEROUS_uncachedSystemPromptSection 到 cch=00000 計費地雷,原來 prompt 工程師現在也要是個會計師。

Anthropic 工程師揭密:Claude Code 的 Prompt Caching 設計哲學 — 整個系統都繞著 cache 轉

Anthropic 的 Claude Code 工程師 Thariq 分享了他們從實戰中學到的 prompt caching 教訓:system prompt 排列順序決定一切、tools 不能加不能刪、model 不能中途換、compaction 要共享 prefix。他們甚至會對 cache hit rate 發 SEV。如果你正在做 agentic 產品,這篇是教科書等級的實戰經驗。

LLM Context Tax 避稅指南:13 招讓你的 AI Agent 帳單少一個零

每個 token 都是錢、都是延遲、過了某個點還會讓你的 AI 變笨 — 這就是 Context Tax 的三重懲罰。Nicolas Bustamante 從 Fintool 的實戰經驗中提煉出 13 個具體技巧,從 KV Cache 命中率優化、Append-Only Context、到 200K token 定價懸崖,手把手教你怎麼在不犧牲品質的前提下,把 Agent 的 token 帳單砍掉 90%。這不是理論文,這是真金白銀的省錢指南。