On October 20, 2026, OpenAI added optional 12-hour prompt caching to the API. Shared instructions, tools, and context can remain reusable across requests hours apart, potentially benefiting agents that work intermittently.
Caching is not permanent memory
This is an input-reuse and cost mechanism, not a promise that a model permanently remembers new information. Reuse still depends on supported models and matching conditions. Asking a later message to reuse material does not prove a cache hit.
Account for writes
Twelve-hour cache writes cost 2.5 times the normal input-token rate. A large context used once may not justify that cost, while repeated reuse may. Calculate with the selected model’s actual rates, cache-hit information, and reuse frequency.
Run a checkable comparison
Choose tasks with stable instructions and documents, then compare ordinary input, cache writes, cache reads, and output charges. Separate changing task details from stable background to investigate misses. Do not promise a fixed saving percentage before measuring total expenditure.
