12-hour prompt caching: when OpenAI’s longer cache saves money

12-hour prompt caching arrived in the OpenAI API on 20 October 2026. You can now keep a shared prompt prefix cached for at least 12 hours instead of the default 30 minutes. The catch is the write price: 2.5× the uncached input rate, double the 1.25× charged for a 30-minute write.

Cached reads cost 0.1× input on most models, and 0.05× on GPT-6.1 Sol, whichever lifetime you choose. So the longer cache pays only when a prefix is reused after the 30-minute window would have expired.

Sources checked on 26 October 2026. Features, prices and availability can change; confirm the linked primary sources before acting.

How 12-hour prompt caching is priced

You pay once to write the prefix and a small rate each time it is read. The write is the full charge for those tokens, not a surcharge on top.

The minimum cacheable prompt is 1,024 tokens on GPT-5.6 and later. The 12-hour option works with the Responses API on those models; earlier models use a separate 24-hour retention setting.

Break-even: when the longer cache pays

It pays from the third request when requests arrive more than 30 minutes apart. When they arrive more often, the default is cheaper.

With gaps over 30 minutes, the default cache expires between requests, so each request writes again at 1.25×. With the 12-hour cache you write once at 2.5× and then read at 0.1×.

A worked example: a 100,000-token prefix on GPT-6.1 Sol, used six times at two-hour intervals. The 30-minute default costs about $1.50 in writes. The 12-hour cache costs $0.50 to write plus five reads at $0.01, about $0.55.

A read within the 30-minute window refreshes it, so steady traffic keeps the default cache warm. A 30-minute read of a 12-hour entry keeps the later expiry; it does not restart the 12-hour clock.

Patterns that benefit, and ones that do not

Use the longer cache for large prefixes reused hours apart. Skip it for constant traffic or one-off prompts.

Put stable content first: system instructions, then tool definitions, then shared knowledge. Variable user input goes last, or nothing after it can be cached.

How to enable it and measure the result

Set the request-level option in the Responses API. A single breakpoint can also override it.

{
  "model": "gpt-6.1-sol",
  "prompt_cache_options": { "ttl": "12h" },
  "input": [ … ]
}

Then read the cached-token count in each response's usage data and compare spend for a week. Our verdict: enable 12-hour prompt caching per workload, only where requests are spaced more than 30 minutes apart and reused at least three times. Our guide to cutting LLM API costs with caching and batching covers the wider picture. For a cost review, see our API development and integration service.

Sources and further reading

• API changelog: 12-hour prompt caching — OpenAI
• Prompt caching guide — OpenAI
• API pricing — OpenAI

Image credits

Explanatory diagrams: Cloudsoft Zone, CC BY-SA 4.0. Featured photograph: PiDatacenters, CC BY-SA 4.0, via Wikimedia Commons.