OpenAI Extends Prompt Caching to 12 Hours

This is infrastructure disguised as a quality-of-life improvement: persistent context reuse could matter more to agent workflows than modest model-speed gains.

• –API developers can choose 12-hour retention for stable instructions, tools, and reference context.
• –The tradeoff is a higher cache-write rate—2.5× ordinary input pricing for 12-hour caching—while cache reads remain discounted.
• –Codex and Work users should see better continuity when resuming long-running tasks or old conversations.
• –Developers still need stable prompt prefixes and consistent tool definitions; changing the prefix can invalidate reuse.
• –Longer retention makes caching more useful for coding agents, where large contexts are repeatedly revisited throughout the day.