The value is not the topic alone, but whether it changes a practical decision today.
Why cache hit rate is the new KPI
As of May 2026, Anthropic's prompt cache has a 5-minute TTL. Send the same prefix and the cached portion is billed at 10% of the input rate. Sending 1M context every call is expensive, but reusing the same prefix inside that 5-minute window crashes the average cost.
The metric that matters in ops is not absolute cost — it's hit rate. Without it, you cannot answer two questions: (1) which call patterns are bleeding the cache, and (2) which workflow rewrites give the best ROI. We started at 14%. After we applied the six patterns below, we hit 87%. Per-call average cost dropped 6×.
The rest of this issue walks through each pattern — (1) split prefixes into a separate prompt component, (2) batch multiple questions on the same prefix, (3) schedule follow-up round-trips inside the 5-minute TTL, (4) auto-refresh the cache just before expiry, (5) stabilize the prompt opening, and (6) log cache hit rate on every call. Each pattern comes with before/after numbers and code.
📖 The remaining six patterns and the measurement data are delivered to Daily Attachment members every morning. The PDF attachment includes code snippets and graphs.
- As of May 2026, Anthropic's prompt cache has a 5-minute TTL. Send the same prefix and the cached portion is billed at 10% of the input rate. Sending 1M context...
- The metric that matters in ops is not absolute cost — it's hit rate. Without it, you cannot answer two questions: (1) which call patterns are bleeding the cache...
- The rest of this issue walks through each pattern — (1) split prefixes into a separate prompt component, (2) batch multiple questions on the same prefix, (3) sc...
→ 한국어 버전 →
- As of May 2026, Anthropic's prompt cache has a 5-minute TTL. Send the same prefix and the cached portion is billed at 10% of the input rate. Sending 1M context...
- The metric that matters in ops is not absolute cost — it's hit rate. Without it, you cannot answer two questions: (1) which call patterns are bleeding the cache...
- The rest of this issue walks through each pattern — (1) split prefixes into a separate prompt component, (2) batch multiple questions on the same prefix, (3) sc...
→ 한국어 버전 →
Take-aways
- As of May 2026, Anthropic's prompt cache has a 5-minute TTL. Send the same prefix and the cached portion is billed at 10% of the input rate. Sending 1M context...
- The metric that matters in ops is not absolute cost — it's hit rate. Without it, you cannot answer two questions: (1) which call patterns are bleeding the cache...
- The rest of this issue walks through each pattern — (1) split prefixes into a separate prompt component, (2) batch multiple questions on the same prefix, (3) sc...
→ 한국어 버전 →
