Semantic Caching vs Prompt Caching: Complementary 73% Cost Reduction for High-Repetition Agent Workloads

Redis Engineering·high signal

Redis engineering documents that semantic caching (storing response embeddings and retrieving on similarity) reduces LLM costs 73% and latency 88% in production workloads with repetitive queries — distinct from prompt prefix caching which only works on identical tokens. A $47K/month LLM bill dropped to $12.7K with Redis LangCache, and AWS chatbot experiments hit 86% cost reduction at >90% accuracy. The correct stack: use prompt caching for shared static prefixes across all calls, then semantic caching to deduplicate semantically equivalent queries at the application layer — the two techniques address different budget centers and compound when combined.