October 23, 2026 · Ahmed · More by Ahmed
OpenAI's New 12-Hour Prompt Cache, and When It Pays Off
OpenAI's Responses API now offers a 12-hour cache lifetime at 2.5x the write cost. Here's the rate math that decides whether it's worth it.
What's New: A 12-Hour Cache Lifetime Option
On October 20, 2026, OpenAI's API changelog announced that the Responses API now supports a 12-hour minimum prompt-cache lifetime for GPT-5.6-and-later models, alongside the existing 30-minute default. Prompt caching already discounts reused input tokens — when a later request shares the same prefix as an earlier one, OpenAI can skip reprocessing it and charge a reduced rate instead. What changed is how long a cached prefix is guaranteed to stick around before it becomes eligible for eviction.
The default lifetime, 30 minutes, carries a cache-write rate of 1.25x the ordinary uncached input-token rate. The new 12-hour option costs 2.5x to write. In both cases, reading a cached prefix back costs 0.1x the ordinary input rate on most GPT-5.6-and-later models (0.05x specifically on GPT-6.1 Sol) — the read discount doesn't change based on which lifetime you chose for the write. These are total write rates, not surcharges layered on top of ordinary input pricing, and choosing "12h" sets a minimum, not a hard deletion deadline: OpenAI may retain entries longer, and the setting doesn't change an org's maximum data-retention policy.
The Rate Math, and Where It Breaks Even
OpenAI's own prompt-caching guide works through the arithmetic for the default lifetime: at the 0.1x read rate, writing a prefix once and reusing it in full once costs 1.35x what processing it twice with no caching would cost, versus 2x uncached. Stretch that to ten requests — one write, nine full reads — and the cached total comes to 2.15x the single-request cost, against 10x if nothing were cached at all.
That 1.35x / 2.15x math is the baseline a 2.5x write has to beat. The guide doesn't publish an equivalent worked example for the 12-hour tier, but the mechanism is the same: a write is a one-time cost, amortized across however many reads land inside the lifetime window before the prefix expires. A 30-minute lifetime amortizes a 1.25x write across whatever reads happen to fall inside half an hour; if your request cadence is slower than that, the prefix is gone before a second use, and you've paid the premium for a single read or none. A 12-hour lifetime gives a slower cadence far more room to accumulate reads against one write — at twice the write cost.
Configuring It, and Where the Edges Are
The setting lives at prompt_cache_options.ttl in the Responses API, with supported values "30m" (default) and "12h"; it applies only to GPT-5.6 and later — earlier models instead use a separate prompt_cache_retention parameter with different semantics. One edge case worth knowing before mixing lifetimes in a pipeline: upgrading a prefix that already has valid 30-minute coverage to 12-hour coverage charges a brand-new 12-hour write for those tokens, with no credit for the 30-minute write already paid. If a given prefix is a candidate for the 12-hour tier, it's worth deciding that up front rather than switching mid-stream.
For a scheduled job — a daily or hourly pass that reuses the same system instructions every run — the deciding factor is simply how far apart requests actually land relative to 30 minutes, and whether the savings from skipping reprocessing justify a write premium that starts out twice as expensive.