Until now, an AI agent remembering your instructions was a bit like a goldfish with a great attention span: brilliant for half an hour, then a blank stare. OpenAI just handed developers a much longer leash, and the people who'll notice first are whoever pays the API bill.

From 30 Minutes to a Full Shift

On October 20, OpenAI added a 12-hour prompt caching option to its Responses API. Developers can now set prompt_cache_options.ttl to "12h" instead of the 30-minute default. That keeps shared instructions, tool definitions and reference context cached for a minimum of 12 hours, so later requests pay the cheaper cached-input rate even when they arrive hours apart.

You can also mix both settings in one request: 12 hours for stable material like system instructions, 30 minutes for fresher conversation turns. The 12-hour option works with GPT-5.6 and later models. Caching kicks in at 1,024 visible input tokens, and OpenAI notes that neither setting is a hard expiration time, since entries may stick around longer.

The Fine Print Has a Price Tag

Here's the catch: writing to the 12-hour cache costs 2.5 times the normal input-token rate, compared with 1.25 times for a 30-minute write. Reads are the payoff, at 0.1 times the normal rate on most supported models and 0.05 times on GPT-6.1 Sol. So the math only works if you actually reuse that cached context. Cache a giant prompt once, never touch it again, and you've just paid a premium for nothing.

The bigger story is what this says about where AI is heading. OpenAI framed the feature around persistent agents that reuse context throughout the day: support bots, internal assistants, content tools that run on schedules instead of one-off chats. For businesses building those, cost structure matters as much as model quality. A well-designed agent with a stable, cache-friendly prompt could get noticeably cheaper to run. A messy one, rewriting its instructions on every call, gets nothing. OpenAI also says signed-in Codex and ChatGPT Work users now get cached context retained indefinitely, a separate consumer-side change that hints at how much the company wants long-running AI work to feel seamless.

Long memory is a feature; using it wisely is a skill. Measure your reuse before you flip the switch.

Building an AI assistant into your site or workflow and wondering whether caching tricks like this will save you money? Our team can help you design an integration that's smart about both answers and invoices.

Source: OpenAI Developer Docs