GPT-6 prompt caching gets a 30-minute reuse window, effort changes that survive the cache, and a miss diagnostic
Cache discounts now apply to eligible shared prefixes reused within 30 minutes, at up to 90% off cached input tokens.3 Two controls matter for agents: reasoning effort can be changed between responses without breaking cache, by appending a configuration_update rather than changing request-level effort; and tools should be masked with allowed_tools or tool_choice: none instead of being removed, so definitions stay stable. A diagnostics endpoint returns the reason for a miss — tools_changed, with the count of tokens that would otherwise have been reusable.
Relevant to: Anyone running a long-lived agent loop on GPT-6 whose input bill is mostly re-sent context.
The miss payload, and what OpenAI claims for it
{
"prompt_cache_diagnostics": {
"type": "cache_miss",
"reason": "tools_changed",
"comparison_reusable_tokens": 5629,
"cache_missed_tokens": 5629
}
}A Prompt Caching Dashboard reports hit rates over time and splits input into cached and uncached tokens.
OpenAI quotes GitHub's Mario Rodriguez saying Copilot cut the share of prompt tokens needing fresh processing by more than 50% across billions of requests against their previous baseline. That is a customer testimonial on a vendor page, not a measurement we can reproduce, and we print it as such.3
Prewarming lets an application process known context — shared instructions, tool definitions, reference material — before the first user request arrives.