OpenAI's API can now keep a prompt cache for 12 hours, at twice the write price of the 30-minute default

OpenAI added a 12-hour prompt cache to its Responses API on October 20, according to its API changelog. On GPT-5.6 and later models, developers set prompt_cache_options.ttl to "12h" instead of the default "30m", and a cached prefix is kept for at least 12 hours, so instructions and reference material reused hours apart still get the cached-input price.

THE PRICE

The write price is what changes. Per OpenAI's prompt caching guide, a 12-hour cache write costs 2.5 times the normal input price, against 1.25 times for the 30-minute default, and those are total rates, not surcharges on top of input. Reads cost the same either way. On GPT-6.1 Sol, at $2 per million input tokens, that works out to $5 per million for a 12-hour write, $2.50 for a 30-minute write and $0.10 for a cached read.

THE MATH

One 12-hour write costs as much as two 30-minute writes. By our arithmetic, the 12-hour setting breaks even when a prefix would otherwise expire and be written twice, and saves money from the third time on. For context reused within the same half hour, the default stays cheaper.

CODEX AND WORK

Thibault Sottiaux of OpenAI wrote on X the same day that for logged-in users of Codex and Work, OpenAI now keeps the cache indefinitely, so returning to an old conversation hours or days later still hits it. That is his statement; OpenAI's docs describe only the API setting.

WHAT WE DO NOT KNOW

Whether the indefinite cache in Codex and Work has practical limits, and whether the 12-hour option will reach the Chat Completions API. Models before GPT-5.6 keep the older prompt_cache_retention setting, which has defaulted to 24 hours since May for organizations without zero data retention.