Context Caching's "saving 90% of inference costs" is an upper limit under specific workloads rather than a universal value; when designing, stable content should be placed first and volatile content later, and attention should be paid to the cache TTL and the minimum cacheable length limit.