OpenAI API adds selectable 12-hour prompt caching

OpenAI says developers can now choose 12-hour prompt caching in its API. Prompt caching lets an application that keeps sending the same material, such as its standing instructions, tool definitions and shared context, pay the lower cached-input rate for that repeated input instead of the full price. The discount only applies while the cache still holds that material, so how long it lasts decides which workloads benefit.

That timing problem is what the announcement targets. OpenAI's example is persistent agents, naming one called dots, which return to the same context throughout the day. With the 12-hour option, the post says, their shared instructions, tools and context can stay cached and keep getting the lower rate even when requests are hours apart. The post does not give the default cache lifetime, the cached-input price, any extra cost for choosing the longer option, or which models support it. It establishes the new choice but not the size of the saving.

It follows an earlier OpenAIDevs post on GPT-6 prompt caching, which raised default cache-hit rates and added diagnostics for spotting what prevented reuse. That change was about how often requests reuse the cache. This one is about how long cached material survives between requests.