OpenAI ships better prompt caching tools for GPT-6 agents

OpenAI detailed an improved prompt caching stack for long-running GPT-6 agents: higher default hit rates, discounts of up to 90% on cached input tokens when an eligible shared prefix is reused within 30 minutes, plus a hit-rate dashboard, miss diagnostics, explicit breakpoints, and reasoning-effort changes that keep the saved prefix.

SOFTWARE desk — long agent runs live or die on reused context. OpenAI is productizing cache hit rates the same way cloud vendors productize cold starts, so the bill tracks new work instead of rereading the same tools every turn.

What the page says is new with the GPT-6 family. OpenAI says it launched an improved prompt caching system that delivers higher cache hit rates by default. A hit rate is the share of input that came from the cache instead of being read fresh. It now gives cache discounts for eligible shared prefixes reused within a 30-minute window. A prefix is the unchanged opening of the prompt, the part two requests have in common. Eligible and shared are the page’s limits. A prefix reused after that window is not described as discounted on this page. The page also says it is introducing tools to monitor cache performance, diagnose misses, and choose how much of a prompt to cache. A miss means the request did not reuse the saved prefix. Higher by default is the company’s claim. This desk did not measure a hit rate.

The monitor is the Prompt Caching Dashboard. The page says it shows how much of an application’s input is served from cache. A developer can track hit rates over time and use an input composition chart to compare cached tokens and uncached tokens. Those views are meant to spot a drop in cache hits and to see how a change in the application changes caching. The link on the page goes to the OpenAI usage screen, with the prompt-caching section selected. This desk did not sign in.

When a miss is unexpected, the page points to a prompt caching diagnostics tool. The tool compares a request with a recent response and names what changed: the model, the tools, the settings, or the input. It also estimates how many tokens were affected, so a team can see the size of the miss. The page prints one example of that report. The type is cache_miss. The reason is tools_changed, which means the tool list changed. comparison_reusable_tokens is 5629. cache_missed_tokens is 5629. 5629 is the example on the page. It is not a customer this desk measured, and it is not a request this desk sent.

Choosing what to cache is an explicit cache breakpoint. A breakpoint is a mark the developer places so OpenAI saves the prompt up to that point and can reuse it later. The page says a refreshed prompt caching guide explains how to use the marks, how long a cached prefix stays eligible, and how changes to tools and inputs affect reuse. The guide, which this desk read as the page the post links, is more specific than the post on price. For GPT-5.6 and later, a cache read costs 0.1 times the ordinary input rate. That is a 90% discount, which matches the post’s “up to 90%.” A cache write, the first time that prefix is saved, costs 1.25 times the ordinary input rate. The guide’s arithmetic: writing a prefix once and fully reusing it once costs 1.35 times the ordinary input, against 2 times if both requests are read fresh. Later reads are where the saving grows. The only lifetime value the guide says it supports is 30 minutes, and that is the default. A cached prefix stays eligible for 30 minutes after the latest write or the latest reuse. Reusing it refreshes that clock without another write charge. OpenAI says it may keep the entry longer than 30 minutes. The shortest prefix the guide says can be cached, for GPT-5.6 and later, is 1,024 visible input tokens. Those prices and the 1,024 figure are the guide’s. They are not a sentence on the launch post. This desk did not receive an invoice.

Changing how hard the model thinks, without throwing away the cache. The post says that on GPT-6 models a developer can change reasoning effort between responses without breaking the cache. Reasoning effort is the setting for how much work the model does before it answers. Raise it for a harder task. Lower it for a routine follow-up. The method on the page is to append a configuration_update and leave the request-level reasoning effort unchanged. Append means add it after the saved opening, not edit the opening. The guide says the same for supported GPT-6 and later models, and it says changing the top-level reasoning effort can rewrite hidden instructions and spoil the saved prefix. One optimization bullet on the guide names GPT-6 Astra for that append. The post says GPT-6 models. This filing does not treat those two sentences as one model name. This desk did not send the update.

Keeping the cache when the tools change. The post says to keep tool definitions, schemas, and ordering stable so earlier context stays reusable. A schema is the shape of the arguments a tool accepts. Ordering is the sequence of those tools in the request. Instead of deleting a tool, the post says to use allowed_tools so only the relevant tools can be called, or to set tool_choice to none when no tool is needed. New instructions go in a new developer message toward the end of the context, so they override older ones without rewriting the opening. The guide repeats that list and adds that tools discovered later can be appended at the end. These lines are OpenAI’s. This desk did not rearrange a tool list.

Prewarming is the wait-time piece. The post says prewarming prepares known context ahead of time so the model can start answering sooner when a request arrives. The example: an application can prewarm shared instructions, tool definitions, or reference material during startup, before the person asks the first question. That moves the processing out of the person’s wait. The post says these controls are optional and sit on top of the engine’s default performance. This desk did not time a startup.

Whose results the page prints, and whose they are. Mario Rodriguez, chief product officer, is quoted on GitHub Copilot. The quote says that over the past several months the share of prompt tokens that needed fresh processing fell by more than 50%, across billions of requests to OpenAI models, relative to the previous baseline, and that the first response got faster. More than 50% means fewer than half as many of those tokens were processed from scratch, if the claim holds. Billions is the quote’s scale. The byline is chief product officer. The quote names GitHub Copilot. The page this desk read does not add a second title. A quote is not a measurement this desk reran. The GPT-6 Sol and Luna page, already filed, cites a GitHub line with the same several-month figure. This page prints it as Rodriguez’s words. It is not a new count of today’s discount.

Arian Hanifi, chief technology officer, said the diagnostics and the dashboard improved cache hit rates by a few percentage points and cut costs by 20%. He said the team now gets alerts when caching breaks and uses Codex agents to find the cause. Explicit breakpoints, he said, cache the stable context and leave frequently changing content at the end, which made it practical to split off a second conversation for background work while reusing nearly all of the shared context. 20% is the quote’s cost cut. A few percentage points is the quote’s hit-rate move. The byline is chief technology officer. The page this desk read does not print an employer. Do not add one. Bin Fan, agent team lead, said that for long-running agents like Manus, reliable caching is fundamental to the economics. Working with OpenAI’s engineering team, they refined where breakpoints sit, combined explicit and automatic caching, and used real requests to find unexpected misses. In less than a week, the quote says, the OpenAI-model cache hit rate went from roughly 85% to consistently above 90%. The quote names Manus. The byline does not print “of Manus” as a title. Eugene Mikhantyev, AI engineer, said that after moving session agents to explicit breakpoints, evaluation hit rates rose from 83% to 91% in under a week. Cache writes fell by roughly two-thirds, and inference costs fell by 36%, on the same workload. Inference is the work of running the model. Two-thirds and 36% are the quote’s. The page does not name that team. Do not name one. All four sets of figures are the speakers’, as OpenAI printed them. This desk did not audit a dashboard.

Plain English for the rest of the card: token = a small chunk of text. input = what you send. cache = a saved opening of a prompt. prefix = the part two requests share at the start. hit rate = the share of input served from the cache. miss = a request that did not reuse that save. breakpoint = a mark that says cache up to here. 30 minutes = how long the post and the guide say an eligible prefix stays reusable after the last write or reuse. up to 90% = the post’s discount ceiling on cached input tokens. 0.1 times = the guide’s read price for GPT-5.6 and later, which is that 90% off. 1.25 times = the guide’s price to write the cache the first time. 1,024 = the guide’s shortest cacheable prefix, in visible input tokens. reasoning effort = how hard the model works before it answers. configuration_update = the item you append to change that effort without rewriting the opening. allowed_tools = a limit on which tools can run, without deleting them. tool_choice none = no tool this turn, with the definitions still present. prewarm = do the saved work before the person is waiting. 5629 = the token count in the page’s example miss, not a customer total. This filing is the 22 Sep caching post. It is not a sticker cut on the model.

PRIMARY here: OpenAI’s 22 Sep 2026 page, “Better prompt caching for GPT-6,” at openai.com/index/better-prompt-caching-for-gpt-6/ — Tier A PRIMARY, the company’s own record. The page did not print an hour. The refreshed prompt caching guide is the document that post links for how breakpoints, lifetime, and tool changes work, also OpenAI’s. The hours-long agents, the up-to-90% discount, the higher default hit rates, the 30-minute window, the dashboard, the diagnostics example with tools_changed and 5629, the breakpoints, the GPT-6 reasoning-effort line, the tool-stability guidance, the prewarm example, and the four quotes are the post’s. The 0.1-times read, the 1.25-times write, the 1.35-times one-reuse arithmetic, the 30-minute lifetime that refreshes on reuse, the note that OpenAI may retain an entry longer, and the 1,024-token minimum are the guide’s. NOT claimed: that this desk measured a hit rate, opened the dashboard, sent the example request, named an employer the bylines do not print, treated “like Manus” as a job title, treated the Sol and Luna page’s GitHub citation as a second study, a stock tip, or investment advice. Distinct from the already-filed gpt-6-sol-luna, openai-third-party-assessments, and digitalocean-managed-agents.