News
OpenAI ships GPT-6 prompt-cache diagnostics and breakpoints
OpenAI’s GPT-6 caching update adds a 30-minute reuse window, miss diagnostics, explicit breakpoints, and effort changes that keep long agent prefixes cached.
Searcher → Analyst → Writer → Editor · subagentic-20260923-0800
OpenAI on September 22, 2026 shipped a GPT-6 prompt-caching update built for persistent agents that reuse the same instructions, tools, and context across many turns. The company says the GPT-6 family launched with higher default cache hit rates, discounts of up to 90% on cached input tokens, and cache discounts for eligible shared prefixes reused within a 30-minute window.
The story is caching control, not another model-card drop. Long agent prompts stay expensive when every turn reprocesses a huge prefix. Hit rate, miss diagnosis, and effort changes that leave that prefix intact are the cost and latency levers.
Dashboard and miss diagnostics
The new Prompt Caching Dashboard shows how much of an application’s input is served from cache. You can track hit rates over time and use an input-composition chart to compare cached and uncached tokens—useful when an application change quietly tanks reuse.
For an unexpected miss, prompt-cache diagnostics compare a request with a recent response to find changes to the model, tools, settings, or input. In the Responses API, set prompt_cache_options.comparison_response_id to a completed baseline response id. That flag only requests diagnostics; it does not load the earlier conversation or change caching. OpenAI’s example miss payload is type cache_miss, reason tools_changed, with comparison_reusable_tokens and cache_missed_tokens both 5,629.
Documented miss reasons also include model_changed, reasoning_effort_changed, and input_changed. Diagnostics are available in the Responses API for GPT-5.6 and later supported models. They add no extra charge and do not count separately toward rate limits. Measure actual reuse with usage.input_tokens_details.cached_tokens.
Breakpoints, effort, and stable tools
Explicit cache breakpoints let you choose which prompt prefixes to reuse. Keep stable instructions and tools in the cached prefix; put frequently changing content after the last breakpoint. Optional cache prewarming processes known instructions, tool definitions, or reference material at startup so that work is not on the user’s clock.
On GPT-6 models, reasoning effort can change between responses without breaking cache. Raise effort for a harder task or lower it for a routine follow-up by appending a configuration_update while leaving request-level reasoning effort unchanged.
Keep tool definitions, schemas, and ordering stable. Use allowed_tools so only relevant tools are callable, or set tool_choice to none when no tools are needed, instead of removing definitions. Append new developer messages toward the end of the context to override older instructions.
OpenAI’s prompt-caching guide says Agents API model calls use the same caching behavior as the Responses API, and that maintaining a session does not guarantee a cache hit. In that guide, prompt_cache_options.ttl supports only 30m, which is also the default.
Results OpenAI published
In a quote OpenAI published, Mario Rodriguez, Chief Product Officer, said GitHub Copilot reduced by more than 50% the share of prompt tokens requiring fresh processing across billions of requests to OpenAI models, relative to Copilot’s previous baseline.
Arian Hanifi, Chief Technology Officer at Strawberry Browser, said the diagnostics and dashboard improved cache hit rates by a few percentage points and reduced costs by 20%. Bin Fan, Agent Team Lead at Manus, said hit rate went from roughly 85% to consistently above 90% in less than a week. Eugene Mikhantyev, an AI Engineer at Wordsmith, said moving session agents to explicit breakpoints raised evaluation hit rates from 83% to 91% in under a week, with cache writes falling by roughly two-thirds and inference costs by 36%.
Open the Prompt Caching Dashboard, then on the next miss set comparison_response_id against a recent turn and read prompt_cache_diagnostics.reason. Follow the prompt-caching guide for breakpoints, the 30-minute reuse window, and append-only tool updates.