GPT-Live delegation almost never reuses prompt cache across sessions: 1 of 1,184

chasef07 · x · 2026-11-21

Developer chasef07 flagged a prompt-caching issue with GPT-Live + Responses delegation to OpenAI: the backend model (gpt-6-luna) sends the same 12.5K-token instructions every call, caching works great within a call (97% hits), but across calls it almost never reuses — only 1 of 1,184 sessions last week.

That means an identical prompt pays a fresh cache write on every call. He asks whether delegated caches are session-scoped by design or a bug, and whether promptcachekey or a cache TTL can be set; others reported the same on the OpenAI forum (the first backend delegation in every new GPT-Live session gets cachedtokens = 0, while later delegations in-session hit over 95%).

Original post →