OpenAI's GPT-6 prompt caching adds explicit cache breakpoints and a cache-miss diagnostics tool
OpenAI Blog·medium signal
OpenAI's caching update for the GPT-6 family lets developers mark cache breakpoints, change reasoning effort mid-conversation without invalidating the cache, and prewarm a cache before a user request arrives. Cached input is discounted up to 90%, shared prefixes stay eligible for 30 minutes, and a new dashboard and diagnostics tool report hit rates and name the cause of each miss, such as a tool change. Long-running agents that rebuild context every turn get the largest savings.