OpenAI extends reusable prefix retention to spaced streams. See when the savings pay off and how to measure privacy, cost and success.
Direct answer
On October 20, 2026, OpenAI added the option to cache prompts for 12 hours in the Responses API. The feature keeps instructions, tools, and shared context available longer for flows whose steps occur hours apart. Writing to the cache costs 2.5 times the normal entry price, so the benefit depends on actual reuse, prefix stability, and comparing write cost, discounted reads, and avoided latency.
Long cache serves asynchronous work
Agents awaiting approval, documents, or external systems can reuse a large context without requiring all steps to happen in the same short window.
Stable prefix determines the hit
Common instructions and tools must come before variable data. Frequent changes to the beginning of the prompt break the reusable identity and reduce the gain.
Writing Price Requires Break-Even Point
Calculate how many reads will occur within 12 hours and compare the total spend for uncached calls. Rare or highly variable flows can be more expensive.
Metric needs to separate attempt and benefit
Record input tokens, cached tokens, call interval, latency and cost per completed task; success rate alone does not prove return.
Nexus Reading
Use the extended cache in demonstrably spaced journeys, with versioned prefixes, minimized data and alerts when the expected savings do not appear.
FAQ
Is the 12 hour cache automatically active?
No. The organization must use the supported option in the API and structure the prompt for reuse.
Does it always reduce costs?
No. Writing costs more and only pays off when the same prefix is reused sufficiently within the window.
What should be monitored?
Tokens written and read from cache, hit rate, latency, cost per result, and prompt version changes.
Essential guides to delve deeper into the decision
This editorial analysis was produced by Nexus from the official sources below, consulted on October 27, 2026. The text is original and interprets practical implications for companies.
- OpenAI Developers — API changelog: date, availability, and scope of the prompt cache for 12 hours.
- OpenAI Developers — Prompt caching: functioning, metrics, prefix structure and practices for achieving cache hits.
Date reported by the main source: October 20, 2026.