Anthropic prompt caching cut our RCA cost by 90%

Short summary

Anthropic's prompt caching cut root-cause-analysis input costs 90% in production by caching static prompt segments at $0.10/million vs. $1.00 base rate. The key is separating cacheable parts (system prompt, retrieval context) from incident-specific data using independent cache_control markers. With proper structuring and bursty call patterns, 60-70% cost savings are achievable within 5-minute TTL windows.

• •Cache read costs 10% of base input rate; break-even point is ~1.25 repeating calls per cached segment within 5-minute TTL
• •Separate system prompt and retrieval context into independent cache markers to prevent cache invalidation across tenants
• •Real production RCA use case shows 70-80% of input tokens are cacheable, reducing total input cost by 60-70% after overhead

Generated with AI, which can make mistakes.