Claude Prompt Cache Complete Guide: 5m / 1h Tier Testing

1. What Is Cache?

Anthropic offers a prompt cache mechanism: the second call with the same prompt gets the input price discounted. See exact discounts in the Claude Opus 4.7 Official L3 Details, which includes three tiers: cache_read 5m / cache_write_5m / cache_write_1h.

2. 3-Tier Pricing (Tested)

• cache_read_5m — re-read the same prompt within 5 minutes, $0.30/M (10% of base price)
• cache_write_5m — write to the 5-minute cache, $3.75/M (25% of base price)
• cache_write_1h — write to the 1-hour cache, $6/M (40% of base price)

3. 6 Scenarios Where You Should Enable Cache

1. Long system prompt (>1k tokens) — high re-read cost, cache is a must
2. Multi-turn conversations — from turn 2 onward, all history is a cache hit
3. Batch tasks (same prompt 100x) — calls 2-100 are all cache hits
4. RAG retrieval — retrieved documents are fixed; after caching, only the new query is billed
5. Few-shot — examples are fixed; after caching, only the new input is billed
6. Agent loop — tool descriptions are fixed; after caching, only the new task is billed

4. 3 Scenarios Where You Should NOT Enable Cache

1. Short prompt (<500 tokens) — cache write cost is actually higher
2. Different prompt every time — 0% cache hit rate = wasted spend
3. Won't re-read within 5m — the 5m cache expires, wasting the write cost

5. Decision Checklist

1. Check the cache tier pricing on Anthropic Official

2. Check Laozhang.ai Claude Details = whether the relay passes through cache pricing

3. Run a test with Claude Detection = verify cache protocol passthrough