01
What changed in GPT-6 prompt caching
OpenAI expanded prompt caching for the GPT-6 family on September 23, 2026, adding a Prompt Caching Dashboard, cache-miss diagnostics, explicit breakpoints and new controls for deciding how much stable context an application should try to reuse. OpenAI says eligible shared prompt prefixes reused within a 30-minute window can receive cache discounts.
The change matters most for agents and multi-turn applications that repeatedly send the same instructions, tool definitions and conversation context. OpenAI says GPT-6 delivers higher cache hit rates by default, while the new tooling makes cache behavior observable and easier to debug.
02
Prompt Caching Dashboard and cache-miss diagnostics
The Prompt Caching Dashboard shows how much application input is served from cache. Developers can track hit rates over time and compare cached with uncached tokens using the input-composition view, making regressions visible after prompt, tool or application changes.
For unexpected misses, OpenAI's diagnostics tool can compare a request with a recent response and identify changes to the model, tools, settings or input that prevented reuse. It also estimates the number of affected tokens so teams can judge the impact before changing an integration.
03
Explicit cache breakpoints and reusable context
Explicit cache breakpoints let developers choose which prompt prefixes to reuse. That is useful when a request mixes stable material such as policies, tool definitions or reference context with frequently changing user messages and tool results.
OpenAI also says GPT-6 can change reasoning effort between responses without breaking cache when applications append a configuration_update while leaving request-level reasoning effort unchanged. The current prompt-caching guide should remain the authority for exact API syntax.
04
The 30-minute window and cache prewarming
OpenAI says cache discounts apply to eligible shared prefixes reused within a 30-minute window. That is an eligibility rule, not a promise that every token in every request remains cached for exactly 30 minutes; prefix matching and current API rules still apply.
GPT-6 also supports cache prewarming, which prepares known shared context before a request arrives. OpenAI gives shared instructions, tool definitions and reference material as examples that can be prepared during application startup to move processing outside the user's wait time.
05
How to improve GPT-6 cache hit rates
Keep stable instructions and tool definitions consistent across related requests, place reusable context before highly dynamic material where current API rules permit it, and measure actual cache-hit behavior rather than assuming reuse is working. When hit rate drops, use diagnostics to isolate the changed model, tool, setting or input.
Do not optimize solely for a high cache percentage. The useful production measures are total application cost, latency and task reliability. OpenAI says cached GPT-6 input can receive discounts of up to 90%, but realized savings depend on workload shape and reuse.
06
Why caching matters for long-running agents
Long-running agents repeatedly carry forward instructions, tools and context. Reusing that stable prefix can reduce repeated computation, latency and input cost while task-specific state continues to change.
OpenAI's announcement includes customer examples of improved cache-hit rates and lower costs, including one report of a 20% cost reduction. These are customer-reported outcomes, not guarantees. Caching improves inference efficiency; it does not make an agent more accurate by itself.
07
FAQ
Does GPT-6 prompt caching last 30 minutes?
OpenAI says eligible shared prefixes reused within a 30-minute window receive cache discounts. Treat that as the current eligibility window rather than a guarantee that every request is cached for exactly 30 minutes.
Why is my GPT-6 prompt cache missing?
Changes to the model, tools, settings or request input can prevent reuse. OpenAI's diagnostics tool is designed to compare requests and identify likely causes.
Does prompt caching reduce output-token cost?
Prompt caching concerns reused input context. Do not assume it discounts generated output tokens unless the current model pricing documentation explicitly says so.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.