United StatesOpenAI’s 12-hour cache costs 2.5 times input price to write.
That doubles the 1.25 times rate for the default 30-minute cache, on GPT-5.6 and later models.
OpenAI’s documentation says 12 hours only sets billing eligibility for cache reads, not guaranteed reuse or lower latency.
A day earlier, DeepSeek open-sourced an inference stack designed to retain cache for 3 days.
OpenAI has not yet announced details of the permanent cache promised to Codex and Work users.
How each outlet framed it
- Zhidx
- reads the 12-hour cache as a billing-threshold change, not longer retention, and calls Codex's 'permanent cache' a usage subsidy rather than real storage
Sources: Zhidx