OpenAI's Responses API can now keep a prompt cache for 12 hours, at twice the usual write price

OpenAI’s API changelog for 20 October adds 12-hour prompt caching to v1/responses. The entry says it “keeps shared instructions, tools, and context cached longer to benefit from lower cached-input pricing, even when requests are hours apart,” and that “12-hour cache writes cost 2.5x the input token price.”

How it works

The prompt caching guide puts the option in prompt_cache_options.ttl, for GPT-5.6 and later in the Responses API:

ttl	Minimum lifetime	Cache-write rate
"30m" (default)	30 minutes	1.25x the uncached input rate
"12h"	12 hours	2.5x the uncached input rate

The guide says these are total write rates, not surcharges on top of normal input. Reads cost the same for both lifetimes: 0.1x the ordinary input rate on most of these models and 0.05x on GPT-6.1 Sol.

A few details from the guide that matter in practice:

• The request-level ttl applies to the implicit breakpoint and to explicit breakpoints without their own setting. prompt_cache_breakpoint.ttl overrides it for a single explicit breakpoint, so one request can hold a long-lived shared prefix and a short-lived session block.
• Each request can create up to four cache writes. The minimum cacheable prompt is 1,024 tokens for GPT-5.6 and later.
• "12h" is a minimum, not a deletion deadline. OpenAI may keep entries longer, and the setting does not change the maximum data-retention policy.
• Unsupported values return 400 invalid_request_error. Older models keep the separate prompt_cache_retention parameter.

When it pays

By our arithmetic from those rates: if your requests arrive more than 30 minutes apart, the default cache has expired each time, so every request pays a fresh 1.25x write. With "12h", the first request pays 2.5x and each later one reads at 0.1x. Two requests cost 2.6x against 2.5x on the default. From the third request on, 12 hours is ahead: 2.7x against 3.75x on the default, or 3x with no caching at all. That suits stable system prompts, tool definitions or reference material that a scheduled agent reuses through the day. For a chat or agent loop that hits the cache every few minutes, the default stays cheaper. Our explainer The Prefix You Keep Paying For covers how to order a prompt so the stable part comes first.