OpenAI added 12-hour prompt caching on October 20, giving developers a longer window to reuse shared instructions, tools and reference material. For AI apps and coding agents that return to the same context hours later, the useful question is how often that context earns back its higher write cost.
What changes in the API
For GPT-5.6 and later in the Responses API, set prompt_cache_options.ttl to "12h"; the default remains "30m". Individual explicit breakpoints can override the request setting through prompt_cache_breakpoint.ttl.
The caching guide specifies a minimum of 1,024 visible input tokens. Reuse depends on an identical rendered prefix and a matching cache boundary. Put stable material first, ahead of changing questions.
Work out the reuse, not just the discount
A 12-hour write costs 2.5 times ordinary input, versus 1.25 times for a 30-minute write. Those are total write rates, not additional surcharges. Cache reads cost the same under either lifetime.
At standard short-context pricing, GPT-6.1 Sol charges $2 per million ordinary input tokens, $0.10 for cached reads, $2.50 for 30-minute writes and $5 for 12-hour writes.
Consider a hypothetical 10,000-token reusable prefix, written once and fully read twice within its coverage window. Its 12-hour cost is $0.052: $0.05 for the write plus two $0.001 reads. Processing that prefix three times without caching would cost $0.06. With only one reuse, however, $0.051 exceeds the $0.04 uncached alternative.
The default cache is the more important comparison. One 30-minute write and two hits cost $0.027. Only if all three uses instead require separate short-cache writes does that total become $0.075. These calculations exclude changing suffixes, output tokens and tool charges; they are not measurements.
Choose the lifetime deliberately
Use the longer option for stable context with predictable repeat demand across longer gaps. Frequent, continuous reuse can favor the cheaper default. Track cached_tokens, cache_write_tokens and the bill before widening deployment.
Upgrading existing 30-minute coverage incurs a new 12-hour write without credit for the earlier write. The selected lifetime also does not guarantee lower latency or establish a deletion deadline.
Before enabling it for sensitive workloads, verify the model and organization’s retention settings against OpenAI’s data controls. The announcement does not establish universal 12-hour compatibility with Zero Data Retention.
Cover: Original MacDigest illustration, typeset in code.
