Guide
Prompt caching explained for context-heavy AI systems
Prompt caching reuses a provider-recognized portion of a request. It rewards stable prefixes; compression changes the request and therefore needs cache-aware policy.
Caching and compression are complementary
Cache stable instructions when they repeat. Compress changing tool or retrieval context when it is oversized or irrelevant.
Watch cache writes and reads
A cache miss or rewritten prefix can erase expected savings. Track provider usage rather than inferring it from prompt length.
Keep unknowns visible
Different providers expose different cache fields and prices. A cost report should not treat unavailable data as zero.