OpenAI API adds 12-hour prompt caching: GPT-5.6+ models can keep shared context cached for half a day, with writes at 2.5x input and reads still at 0.1x

Background

Long-running agents resend the same instructions, tool definitions and reference material on every call. On GPT-5.6 and later the default cache only guarantees about 30 minutes, so a request an hour or two later rewrites the prefix at 1.25x. OpenAI Developers named always-on agents like dots as the target.

How it works

The Responses API now accepts prompt_cache_options.ttl of "30m" or "12h", and prompt_cache_breakpoint.ttl can override it per breakpoint. The docs suggest putting all-day content first with a 12-hour breakpoint and session context at 30 minutes; a request can create up to 4 writes. Each token is billed once at the longest lifetime covering it, and upgrading 30-minute coverage to 12 hours is a fresh write with no credit.

Numbers

This is a billing feature, so there are no benchmarks. Vendor-stated rates: 12-hour writes at 2.5x ordinary input, 30-minute at 1.25x; reads at 0.1x on most models and 0.05x on GPT-6.1 Sol. Our estimate: one prefix reused hourly six times a day costs about 7.5x input with 30-minute caching versus about 3.0x with 12-hour caching; for reuse within minutes, 30 minutes stays cheaper.

Getting started

Put stable instructions in a developer-message input_text block with a 12-hour breakpoint and keep changing content after the last breakpoint so it isn't written. Track ttl_12h_tokens under cache_write_tokens_details and cached_tokens to confirm savings. 12 hours is a minimum lifetime, not a hit guarantee; GPT-5.5 and earlier still use prompt_cache_retention.