Discounted pricing even for requests hours apart
OpenAI announced on its official developer account @OpenAIDevs on Oct. 20 (Oct. 21 Japan time) that it has added an option in its API to keep "prompt caching" for 12 hours. Prompt caching temporarily stores identical input sent each time and reads it back at a lower price on subsequent requests.
Until now, the standard cache duration was 30 minutes. For always-on agents that resend large inputs, such as fixed instructions and tool definitions, every few hours, the cache expired in between and the discounted pricing could not be used. OpenAI has in mind uses like its own always-on agent "dots." It can be used by setting the cache duration to "12h" in the API's "Responses API."
Writes at 2.5 times, reads at one-tenth
With the 12-hour cache, writes are priced at 2.5 times standard input. Writes to the 30-minute cache are 1.25 times. Reading cached input costs one-tenth of standard in both cases, and one-twentieth for "GPT-6.1 Sol."
| Pricing type | Multiple of standard input price |
|---|---|
| Write to 30-minute cache | 1.25 |
| Write to 12-hour cache | 2.5 |
| Read of cached input | 0.1 |
| Read (GPT-6.1 Sol) | 0.05 |
Savings depend on the amount of reuse
Because writes are more expensive, whether costs go down depends on how many times the same input is read back. If there are few reads within the 12 hours, it ends up costing more. U.S.-based Runtime Wire noted that simply keeping an agent running does not save money, and that decisions should be based on measuring the actual amount of reuse.
"12 hours" is not a deletion deadline
The 12 hours is the minimum storage period, not a deadline at which data is deleted. OpenAI may store data for longer, and the company said the option should not be relied on as a means of managing data retention periods.
OpenAI is also improving caching in its own services. On Oct. 16, it explained that it had extended cache retention in its programming assistant "Codex" so that more work can be done within the same usage limits.
