OpenAI adds 12-hour prompt cache for GPT-5.6+ API

Day 16 sits inside OpenAI's 28-day shipping cadence and follows a run of changes aimed at long-context work. Day 7 raised GPT-6.1 Sol's standard context from 272K to 500K tokens, and Day 12 improved context reuse and routing between sub-agents. Day 16 shifts the focus from capacity to persistence: keeping a reusable cache alive across hours-long gaps, both through an explicit API setting and through a product-level guarantee for logged-in Codex and Work users.

The two halves of the announcement are governed by different rules. The API option is a TTL choice that offers reuse opportunities but does not promise a hit on any individual request, while the Codex / Work guarantee describes the experience of returning to an old thread. Existing users should also note that the new prompt_cache_options.ttl parameter is not the same as the older prompt_cache_retention used by earlier models, so a "12h" value cannot simply be passed to the old field.

The practical effect depends on usage patterns. A long shared prefix reused at wide intervals can make the higher write cost worthwhile, but frequent use within 30 minutes may favor the default. Cost measurement still hinges on the usage breakdown, since cached tokens continue to count toward the TPM rate limit.