OpenAI Adds 12-Hour Prompt Caching to the Responses API
OpenAI added an optional 12-hour prompt-cache lifetime to the Responses API for GPT-5.6 and later. Developers can keep matching prompt prefixes eligible for cached-input pricing across longer gaps by setting `prompt_cache_options.ttl` to `12h`; the default remains 30 minutes. The longer cache write costs more, so the tradeoff depends on reuse.
OpenAI has added an optional 12-hour prompt-cache lifetime to the Responses API for GPT-5.6 and later. The feature is aimed at stable instructions, tools, and reference material that an application may reuse after gaps longer than the default 30-minute minimum lifetime. Developers can configure the request-level prompt_cache_options.ttl as 12h; the setting applies to the implicit cache breakpoint and to explicit breakpoints that do not set their own lifetime (OpenAI API changelog, prompt caching guide).
What the longer lifetime changes
Prompt caching can reduce the price of input tokens when a request reuses an eligible prefix. With the new option, shared material can remain eligible for cached-input pricing across longer intervals, rather than relying on the default 30-minute minimum cache lifetime. The 12-hour value is a minimum lifetime, not a deadline for deleting a cache entry, and it does not change OpenAI's maximum data-retention policy (OpenAI prompt caching guide).
The option is available through the Responses API on GPT-5.6 and later. The guide recommends using a longer lifetime for stable instructions or reference material reused across longer gaps, while keeping short-lived conversation or agent context at 30 minutes. An explicit breakpoint can set its own TTL when a request needs different lifetimes for different sections of a prompt (OpenAI prompt caching guide).
The pricing tradeoff
OpenAI prices a 12-hour cache write at 2.5 times the ordinary uncached input-token rate. Cache reads keep the model's discounted cache-read rate, which the documentation lists as 0.1 times the ordinary input rate for most GPT-5.6-and-later models and 0.05 times for GPT-6.1 Sol. The rates are model-specific, so developers should check the current API pricing for the model they use (OpenAI prompt caching guide).
That higher write price means the longer TTL is not automatically cheaper for every workload. OpenAI says the 2.5-times rate is the total cache-write rate, not an extra surcharge on top of ordinary input pricing. It is a tradeoff to evaluate when a long, stable prefix is reused hours apart; the result depends on how often the prefix is written and then read (OpenAI prompt caching guide).
The TTL alone does not promise a cache hit or lower latency. For GPT-5.6 and later, a cacheable prefix must contain at least 1,024 input tokens, and hidden OpenAI-provided system tokens do not count toward that minimum. Cache entries reside on individual machines and are not shared across organizations or regional-processing boundaries, so reuse still depends on an eligible matching prefix and the request reaching a machine with a matching, unexpired entry (OpenAI prompt caching guide).
What developers should check
The 12-hour setting is documented for the Responses API on GPT-5.6 and later. Earlier models use the separate prompt_cache_retention setting, with different supported values; that older parameter is not the same as the new 12-hour TTL. Teams should also keep prompt prefixes stable and compare actual cached reads and writes in their usage data before applying the longer lifetime broadly (OpenAI prompt caching guide).
For prompts that combine long-lived shared instructions with short-lived session context, explicit breakpoints allow those sections to use different lifetimes. This makes the option most relevant to applications that can separate durable context from conversation-specific material rather than extending every part of a prompt by default (OpenAI prompt caching guide).
Sources
Frequently Asked Questions
What changed in OpenAI prompt caching?
The Responses API now supports an optional 12-hour minimum cache lifetime for GPT-5.6 and later. Developers can set `prompt_cache_options.ttl` to `12h`; when no TTL is set, the documented default is 30 minutes.
How is 12-hour prompt caching priced?
OpenAI documents 12-hour cache writes at 2.5 times the ordinary uncached input-token rate. Cache reads retain the model's discounted cache-read rate, which varies by model.
Does a longer TTL guarantee a cache hit?
No. It establishes a minimum lifetime for cache eligibility, but reuse still depends on an eligible matching prefix and routing to a machine with the cached entry. For GPT-5.6 and later, the prefix must contain at least 1,024 input tokens; hidden OpenAI system tokens do not count. Cache entries are not shared across organizations or regional-processing boundaries.