OpenAI adds 12-hour prompt caching, at 2.5x the write cost
A new cache lifetime option for the Responses API keeps shared instructions and tools cached across requests that are hours apart, not just minutes.
OpenAI added a 12-hour prompt cache option to the Responses API on 20 October. Until now the default minimum cache lifetime was 30 minutes, with cache writes costing 1.25x the uncached input-token rate. The new 12h setting keeps shared instructions, tools and context cached for up to 12 hours, but the write itself costs 2.5x the uncached input rate, double the standard premium.
Cache reads are priced the same regardless of which lifetime you pick: 0.1x the ordinary input rate for most GPT-5.6-and-later models, 0.05x for GPT-6.1 Sol. Only the write cost and how long the cache survives change.
Why it matters: the 30-minute default is fine for anything that hits the API continuously, but it expires before the next call in scheduled jobs, batch pipelines or any agent that only runs every few hours. Those workloads were paying full input price on every run because the cache had already gone cold. Paying a higher write cost once to hold a large system prompt or tool definition set for 12 hours is cheaper than re-processing it uncached each time.
"prompt_cache_options": { "mode": "explicit", "ttl": "12h" }The caveat: for anything that already calls the API within 30 minutes, the standard 1.25x write rate stays the better deal. The 12-hour option only pays off when requests are genuinely spaced out.