OpenAI has announced a new 12-hour tier for prompt caching in its API. An agent can reuse the same system instructions, tool definitions and context throughout an entire day, and even if two requests are hours apart, the cached input is still billed at the low cached rate. Public reports say this was day 16 of OpenAI's 28-day run of consecutive updates.
A change for OpenAI's own products was announced at the same time. Tibo, who leads OpenAI Codex, said on social media that signed-in users of Codex and Work get permanent caching: returning to an old conversation, or pausing for a few hours mid-chat, carries no extra cost when you pick it back up, and a cache hit is guaranteed. OpenAI's developer documentation does not cover this part, so how it is implemented cannot be verified for now.
What the documentation says
According to OpenAI's developer documentation, the 12-hour tier is available only in the Responses API and applies to GPT-5.6 and later models. It is set through prompt_cache_options.ttl, whose options are the default 30 minutes and 12 hours.
The rates are spelled out clearly:
- On the default 30-minute tier, cache writes are billed at 1.25x the standard input price;
- On the 12-hour tier, writes are billed at 2.5x;
- Reads after a cache hit cost 0.1x the standard input price for most models, and as little as 0.05x for GPT-6.1 Sol.
The documentation stresses that 12 hours is a "minimum time to live": it does not set a deletion deadline and does not change OpenAI's maximum data retention policy. Older models still use the earlier parameter, prompt_cache_retention: the in-memory tier generally expires after 5 to 10 minutes of inactivity and lasts at most an hour, while the 24-hour tier keeps data for up to a day.
A back-of-the-envelope calculation
Take a common scenario: an agent attaches a fixed 100,000-token prefix to every request and is woken every 2 hours across a 12-hour day, for 6 calls in all. Below, billing 100,000 tokens once at the standard input price counts as 1 unit.
| Plan | Calculation | Total |
|---|---|---|
| Default 30-minute tier | Cache has expired every time, so 6 writes x 1.25 | 7.5 units |
| 12-hour tier | 1 write x 2.5 + 5 reads x 0.1 | 3.0 units |
That cuts the cost of this prefix by roughly 60%. Conversely, if requests are dense enough that another one always arrives within 30 minutes, the default tier hits every time, and the extra 1.25x write fee on the 12-hour tier is wasted. As a rough guide, the new tier pays for itself once there are two or more gaps longer than half an hour within a 12-hour window.
How it compares with Anthropic
Anthropic's Claude API offers two cache tiers, 5 minutes and 1 hour, and the 1-hour tier charges 2x the base input price for writes. By stretching the lifetime to 12 hours and setting the write multiple at 2.5x, OpenAI makes the retention cost per unit of time clearly lower. The target customers are just as clear: scheduled jobs, long-running coding agents, and assistants that stay open during the day but are only called every few hours.
The "permanent caching" for Codex and Work follows a different logic. It is not open to API customers; OpenAI absorbs the storage cost itself, and in return subscribers are no longer slowed by cold starts in long conversations. OpenAI has not given a timetable for when API customers might get a longer tier.
Sources: OpenAI developer documentation (prompt caching guide), the OpenAI Developers official account, CocoLoop, Anthropic pricing page. Rate multiples follow the figures listed in OpenAI's documentation; the cost comparison is an editorial estimate.