Prompt caching: 12-hour lifetime (OpenAI API docs)

Summary

On 2026-10-20 OpenAI added a 12-hour prompt cache lifetime to the Responses API, according to its API changelog. Setting prompt_cache_options.ttl to "12h" keeps a cached prefix for at least 12 hours instead of the default 30 minutes. Cache writes cost more, and cache reads cost the same.

Key points

• Enable it: set prompt_cache_options: {ttl: "12h"} on a Responses API request. The default is "30m". You can override the lifetime per breakpoint with prompt_cache_breakpoint.ttl. Other values return 400 invalid_request_error.
• Models: GPT-5.6 and later, Responses API only.
• Prices (as of 2026-10-21): a 12-hour write costs 2.5x the uncached input rate, compared with 1.25x for 30 minutes. These are total rates, not surcharges. Reads cost 0.1x input on most models and 0.05x on GPT-6.1 Sol, whatever the lifetime.
• Upgrading cached tokens from 30 minutes to 12 hours bills a new 12-hour write, with no credit for the earlier write. A 30-minute read on a 12-hour entry does not restart the 12-hour clock.
• The minimum cacheable prefix is 1,024 visible input tokens.
• OpenAI suggests "12h" for stable instructions or reference material reused across long gaps, and "30m" for context reused frequently.

Why it matters

Workloads with long idle gaps, such as daily batch jobs, sporadic agents, or support bots with quiet hours, can now keep a large prefix warm across a working day. The cache pays for itself after about two reads at 2.5x write cost and 0.1x read cost, compared with resending uncached. Anthropic's longest cache lifetime is 1 hour.