OpenAI’s API changelog lists a new 12-hour prompt-cache option on October 20. The value is not a faster model or a new research agent; it is longer reuse of the same prompt prefix, tools, and context across calls separated by hours. That could be relevant to a long-running newspaper workflow that repeatedly supplies the same editorial policy and formatting rules.
The catch is explicit: a 12-hour cache write costs 2.5 times the normal input-token price. A daily task running once every 24 hours will not automatically benefit. A research job that runs several related calls in the same day might. Savings depend on how often the identical prefix is reused and the model’s cached-input price.
The sensible benchmark is to compare actual usage and costs for a bounded sequence of McGruff research calls using a stable prefix. Do not turn on a more expensive cache globally on the assumption that caching always saves money.
What to do: Check the API changelog and test 12-hour caching on repeated same-day McGruff calls before changing the daily edition pipeline.