OpenAI’s September 22 caching update adds a dashboard, cache-miss diagnostics and explicit breakpoints for GPT-6 applications. The company says eligible shared prompt prefixes reused within a 30-minute window receive cache discounts, with cached input-token discounts of up to 90%.
Developers can inspect reuse, compare requests and choose cache boundaries. The announcement also describes prewarming context and preserving cache when adjusting reasoning or tool availability. Savings depend on eligible reused input and application behavior, rather than applying uniformly to every request. Official announcement.