OpenAI's rebuilt prompt cache holds a GPT-6 prefix for 30 minutes, reads it at 0.1x, and now names each miss (2026)

OpenAI rebuilt prompt caching for its GPT-6 generation and published the details on September 22. The cache runs by default. A reused prefix stays eligible for 30 minutes after its most recent write or read, writes cost 1.25 times the ordinary input rate and reads 0.1 times, and nothing shorter than 1,024 tokens is cacheable at all. 1 2

Cached input already carried a discount of up to 90%. What it adds, for GPT-5.6 and later models on the Responses API, is the tooling to see how often you earn it. 1 2

Signal	Confirmed detail	What it means for you
What shipped	Caching rebuilt for GPT-5.6 and later, on by default, with a dashboard and diagnostics. 1	Nothing to switch on; the defaults moved.
What it costs	Reads at 0.1x the uncached input rate, writes at 1.25x. 2	A prefix pays off at its second reuse.
What breaks a hit	A changed model, tool list, tier, schema, effort or verbosity. 2	Tool edits are the usual culprit.
Where to look	Prompt caching in the platform's Usage page; a comparison request names the miss reason. 1	Measure before restructuring anything.

What you can control

Breakpoints. prompt_cache_options.mode: "explicit" marks the blocks to reuse, up to four writes per request; mark none and nothing is written to cache. 2
Reasoning effort, mid-session. On GPT-6 models, append a configuration_update item to change effort between turns without touching request-level reasoning.effort. 1 2
Tool changes. Keep definitions, schemas and their order stable; allowed_tools narrows what is callable, and tool_choice: "none" disables tools without deleting them. 1
Prewarming. Build shared instructions or reference material at startup, before the first user request. 1

What it costs, and what breaks a hit

One write plus one full read costs 1.35 times what that prefix costs uncached, against 2 times to process it twice; over ten requests, one write and nine reads come to 2.15 times against 10. 2 Cached states sit on individual machines, so a request routed away from the one holding the entry pays the full rate, and traffic above 15 requests a minute can be routed away. 2

The dashboard reports your hit rate. When a hit does not arrive, the diagnostics tool compares your request with an earlier response you name and returns one classified reason — changed tools, model, tier, output schema, effort, verbosity, cache key, compacted context or earlier input — with an estimate of the tokens lost. It reports only the first cause, returns unavailable when it cannot classify a miss, costs nothing extra, and works under Zero Data Retention. 3

OpenAI's Prompt Caching Dashboard showing a 41.7% hit rate, 0.91 cache reads per write, and cache-read, cache-write and uncached token volumes — OpenAI's Prompt Caching Dashboard, as published with the announcement: hit rate, reads per write, and the cached, cache-written and uncached input split for the previous month, filtered to two projects. The figures are OpenAI's own. 1

What it produced in production

Four customer accounts appear in the note, all self-reported. Mario Rodriguez, chief product officer, reports GitHub Copilot cutting the share of prompt tokens needing fresh processing by more than 50% across billions of requests; Manus went from about 85% to above 90% within a week; Wordsmith reports 83% to 91% with a 36% cut in inference costs; Strawberry Browser reports about 20% lower cost. 1

Why it matters

For an agent that runs for hours on the same instructions and tool definitions, the cached-input discount is where the published price lands, and this release changes the odds of earning it. 1 The 30-minute window and the machine-local cache hand that to steady traffic whose opening stays stable — long sessions, shared system prompts, forked background tasks — and the dashboard is the cheapest way to learn which kind you have. 2