GPT-6 prompt caching gets diagnostics and explicit breakpoints

OpenAI has updated prompt caching for GPT-6 with higher default hit rates, a dashboard for monitoring reuse, and diagnostics that explain why a request missed the cache. Developers can also place explicit breakpoints in prompt prefixes.

The changes target agents that make many related API calls and repeatedly send the same instructions, tools, and context. OpenAI says cached input can receive discounts of up to 90%, making cache behavior a meaningful part of production agent costs.

The part of an agent bill you can design

The caching guide says reuse depends on an exact shared prefix and compatible request settings. For GPT-5.6 and later, at least 1,024 eligible input tokens are needed. A stable block of instructions and tool definitions can be cached; frequently changing context belongs after it. Explicit breakpoints let an application choose where a reusable prefix ends.

There is a cost tradeoff. OpenAI documents cache writes at 1.25 times the uncached input rate and reads at 0.1 times that rate for these models. If a prefix will only be used once, writing it is no bargain. Reused across a long agent session, it can be. The diagnostics guide lets a developer compare requests and inspect reasons such as changed tools or settings when an expected hit fails.

Measure before celebrating

The announcement includes customer savings examples, but those are specific workloads. Track cache hit rate, cache-write tokens, latency, and total cost across representative conversations. A session ID by itself does not guarantee a hit. If the prompt structure changes every turn, the cache cannot rescue an unstable architecture.

Sources and further reading

• Original publication
• OpenAI API: prompt caching guide
• OpenAI API: cache diagnostics