OpenAI adds a 12-hour prompt-cache option

OpenAI's October 20 API update adds optional 12-hour prompt caching. Shared instructions, tool definitions and context can remain eligible for cached-input pricing even when requests arrive hours apart.

Cache writes for this option cost 2.5 times the input-token price. That makes reuse frequency important: keeping context longer is useful only when later requests recover enough of that extra cost.

For agents that revisit the same project throughout a day, the relevant measurement is total task spending. A lower price on cache reads can coexist with a more expensive workflow if little of the stored prefix is reused.