Prompt-cache: what you already save, and what turning it on would save · 0 credits

Answers, on YOUR real traffic and without spending anything, whether explicit prompt caching is worth enabling on this engine. Two figures, deliberately kept apart: measured is what already happened, read from telemetry; projection is an UPPER BOUND of what explicit caching would have done — provider caches are best effort, Anthropic counts the TTL from the start of the request, and past traffic does not predict a changed pattern. measured: null means NOT MEASURED (no calls in the window), which is not the same as zero. applies:false on OpenAI-shaped providers is not a limitation: those cache by prefix match automatically, so there is nothing to switch on and measured already shows the saving. Traffic is attributed by firm + kind + model (telemetry stores the model, not the config id); when another engine of yours shares both, attribution_note says so. No prompt text, no credentials, no fingerprints: counters only. Query: dias (1-90, default 30) and ttl (5m | 1h). Scope: mcp | * (NOT cases).