OpenAI 12-hour prompt caching: 2.5x writes, do the math

OpenAI just stretched its prompt cache from half an hour to half a day. As of October 20, the Responses API supports 12-hour prompt caching, so an agent that touches the same system prompt and tool definitions a few times a day can still get cached-input pricing. The catch is on the write side: a 12-hour cache write costs 2.5x the normal input rate, double the 30-minute option. That's a good deal for some workloads and a tax on others. The difference is arithmetic.

What actually happened

The OpenAI API changelog entry for October 20 adds 12-hour caching to v1/responses, aimed at persistent agents whose requests arrive hours apart. The prompt caching guide has the details for GPT-5.6 and later:

• Set it with prompt_cache_options.ttl: "30m" (default) or "12h". You can also set ttl on an individual prompt_cache_breakpoint.
• Write cost: 1.25x the uncached input rate for 30 minutes, 2.5x for 12 hours. Those are total rates, not surcharges.
• Read cost: 0.1x the input rate regardless of TTL, or 0.05x on GPT-6.1 Sol.
• Minimum: 1,024 visible input tokens before a prompt is cacheable.
• Unsupported values return a 400 invalid_request_error. The TTL is a minimum, not a deletion deadline.

Why 12-hour caching matters for your business

Run the break-even before you change anything. Take a cached prefix and count how often it gets reused when the gaps between requests are longer than 30 minutes. With the 30-minute TTL, every cold request rewrites the cache at 1.25x. With 12 hours you pay 2.5x once, then 0.1x for each reuse.

• Two requests, three hours apart: 30m costs 2.5x (two writes). 12h costs 2.6x. A wash.
• Three requests: 3.75x versus 2.7x. 12h wins.
• Eight requests spread over a workday: 10x versus 3.2x.

If the prefix is used once and abandoned, 12 hours just doubles your write cost.

The workloads that win are the ones that idle. A nightly reconciliation agent, an inbox triage bot that wakes every hour, a sales assistant that a rep opens four times a day. These used to come back cold after a lunch break. Now they don't have to.

Hot loops should stay on 30 minutes. If your agent fires every few seconds, the 30-minute cache never goes cold, and paying 2.5x for a longer TTL gets you nothing.

Prefix discipline still decides everything. The cache matches an exact prefix. A timestamp or a reordered tool list near the top and you pay the write without ever getting the read.

Sources: OpenAI API changelog, OpenAI prompt caching guide.