OpenAI has introduced a dashboard and request-level diagnostics for GPT-6 prompt caching, giving API developers a way to track reuse of repeated instructions and conversation history and investigate cache misses. The September 22 update also adds controls for choosing which part of a prompt to cache, a practical issue for agents that make many related requests.
Prompt caching saves the beginning of one request so a later request can reuse the same text and tool definitions. On GPT-5.6 and later models, that prefix must contain at least 1,024 visible input tokens. For GPT-6, it remains eligible for reuse for at least 30 minutes after the latest write or reuse, according to OpenAI's API guide.
Diagnostics compare requests to explain a miss
The new Prompt Caching Dashboard charts the share of input tokens served from cache and separates cached from uncached input. Request-level diagnostics can compare a response with a recent one and report a reason for missed reuse. OpenAI's example identifies a changed tool definition as the cause and estimates how many otherwise reusable tokens were missed.
An agent can repeat its instructions yet miss the cache because other parts of the request have changed. The diagnostics identify a possible cause; the response's cached-token count shows how much was actually reused.
Cache savings depend on what repeats
OpenAI also documents explicit cache breakpoints, which let an application mark a stable prompt prefix for reuse while leaving frequently changing material after it. For GPT-5.6 and later, writing eligible context to cache costs 1.25 times the ordinary input-token rate; reading a cached prefix costs one-tenth of that rate. Those rates apply to the relevant input tokens, not to an entire API response.
Savings depend on how much of a request repeats and how often it is reused before the cache expires. Writing a prefix costs more than processing the same input once, while subsequent reads cost less. Responses include separate counts for cached reads and cache writes, allowing developers to calculate the cost for their own traffic.
