OpenAI upgrades GPT-6 prompt caching with higher hit rates, new diagnostics

OpenAI has upgraded prompt caching in the GPT-6 API, enabling higher cache-hit rates by default, cached input token discounts of up to 90%, a new dashboard exposing cache-hit rates and volumes of cached versus uncached tokens, and new features including diagnostics, breakpoints, prewarming, and cache-preserving reasoning changes. The improvements target long-running agents that frequently resend instructions, tool definitions, and earlier context across successive API calls. Achieving high cache hit rates still depends on application design, as developers must structure agents so stable instructions and tools remain reusable.