A real-world test shows how Anthropic prompt caching cut one schema-heavy BI agent’s cost by 57% and latency by 14%.