OpenAI adds 12-hour prompt caching; cache writes cost 2.5x input

OpenAI added a 12-hour option for prompt caching in its Responses API on October 20, 2026, according to its API changelog. Cached instructions, tools and context can now stay cached for at least 12 hours instead of the default 30 minutes, so requests that arrive hours apart can still get the lower cached-input rate. The catch is the write price: “12-hour cache writes cost 2.5x the input token price,” the changelog says.

What changed

The prompt caching guide says 12-hour lifetimes work with GPT-5.6 and later models in the Responses API. Developers set prompt_cache_options.ttl to "12h", and can override it per breakpoint with prompt_cache_breakpoint.ttl. Earlier models keep using the older prompt_cache_retention parameter. The minimum cacheable prefix for GPT-5.6 and later is 1,024 tokens.

What it costs

OpenAI prices cache writes as a multiple of a model’s normal input rate. The guide says these are total prices, “not surcharges on top of ordinary input.”

Cache lifetime	Write price	Read price
30 minutes (default)	1.25x input	Same for both
12 hours ("12h")	2.5x input	Same for both

Reads cost 0.1x the input rate on most of these models and 0.05x on GPT-6.1 Sol, the guide says. Upgrading an existing 30-minute entry to 12 hours triggers a new 12-hour write charge, with no credit for the earlier write.

Who it affects

Apps that reuse a long system prompt, tool list or document across requests spread through the day, such as support bots or agents with quiet periods. A 12-hour write costs twice a default write, so it pays off when the cache would otherwise expire and be rewritten more than once between requests. OpenAI notes that a minimum lifetime does not guarantee every request reuses the cache or runs faster, and that “12h” sets no deletion deadline; cached data is kept for at most 24 hours.

To compare what cached and uncached prompts cost across providers, try our AI token cost calculator or the AI models comparison.