OpenAI added 12-hour prompt caching on October 20, 2026, allowing shared instructions, tools, and context to remain cached for requests hours apart. This catch-up report covers that announcement. The business question is whether repeated agent instructions earn back the higher cache-write cost.
What changed in the Responses API
For GPT-5.6 and later, set prompt_cache_options.ttl to "12h" in the Responses API; the default remains "30m". A content-block prompt_cache_breakpoint.ttl can override the request default. Earlier models use the separate prompt_cache_retention parameter; this is not a universal switch for every OpenAI model.
Our practical recommendation: consider longer coverage for a stable operating policy reused by morning, afternoon, and evening agent jobs—not for each job’s changing customer details.
The higher write cost changes the calculation
OpenAI documents total write rates of 2.5× ordinary input for 12 hours, versus 1.25× for 30 minutes; these are not additional surcharges. Reads cost 0.1× on most applicable models and 0.05× on GPT-6.1 Sol, regardless of lifetime.
Our calculation: normalize processing the shared prefix once without caching to one cost unit. At a 0.1× read rate, one 12-hour write plus one full read costs 2.6 units, more than two uncached passes. One write plus two full reads costs 2.7 units, less than three uncached passes. These ratios follow OpenAI’s published pricing relationships; they are not dollar quotes.
That conditional break-even assumes an unchanged prefix, successful full reads, and no rewrites. It does not establish that 12 hours beats 30 minutes: if shorter coverage keeps yielding hits, its cheaper initial write wins. Output, tools, and other workflow costs also remain outside this prefix-only calculation.
Limits to check before changing agent jobs
The documented minimum prefix is 1,024 visible input tokens. A 12-hour lifetime is neither a guaranteed cache hit nor a deletion deadline. Entries cannot be reused across organizations or regional processing boundaries; upgrading valid 30-minute coverage incurs a new 12-hour write without credit.
For rollout, we recommend testing the same scheduled jobs under both lifetimes. Compare actual cached tokens, writes by lifetime, and total cost—not simply the requested TTL. Keep longer coverage only where observed reuse justifies it; the October 20 announcement creates an option, not automatic savings.