OpenAI adds 12-hour prompt cache and 'eternal' cache for Codex to API

On October 20, 2026, OpenAI added the prompt_cache_options.ttl = 12h parameter to the Responses API, enabling 12-hour prompt caching with writes priced at 2.5x input tokens and reads at 0.1x, while announcing indefinite cache storage for Codex and ChatGPT Work users.

Compiled by Sergey KostenchukPublished 2026-10-21Updated 2026-10-21

2026-10-21 Coding OpenAI

💻 OpenAI adds 12-hour prompt cache to API; Codex gets 'eternal' cache

The Responses API now features prompt_cache_options.ttl = 12h: the common prefix—instructions, tools, context—is cached for up to 12 hours instead of 30 minutes. Writes cost 2.5x the input price ($5.00/1M for GPT-6.1 Sol vs. $2.00), and reads cost 0.1x. Available for GPT-5.6 and newer.

🌍 Agent pauses longer than 30 minutes previously reset the cache. The 12-hour cache saves up to ~75% on input tokens: according to CellCog calculations, an agent with a 100k token prefix called once an hour costs $0.61/day instead of $2.40. Anthropic and Google are under pressure.

👤 Simply setting prompt_cache_options.ttl = 12h is sufficient—savings begin with the third use of the prefix within 12 hours; upgrading from the 30-minute cache is billed as a new write. In Codex and ChatGPT Work, old conversations resume instantly and without extra charge.

Source 1: https://developers.openai.com/api/docs/changelog Source 2: https://x.com/OpenAIDevs/status/2112603916076781778

Sources