OpenAI 12-hour prompt caching: when it saves money, and when it doesn't
Ship ItChecked against the official source
Source checked 2 Nov 2026 · Published 2 Nov 2026 · Last reviewed 2 Nov 2026 · By Ctrl+Z Therapy (how we test)
The short answer
OpenAI's prompt caching is enabled by default for supported models, and reused tokens are billed at a cheaper cached-input rate. For GPT-5.6 and later, cache writes cost 1.25x the uncached input rate for the default 30-minute lifetime. Since October 20, 2026, the Responses API also offers a 12-hour lifetime, with writes at 2.5x the input token price. Cache reads cost 0.1x on most of these models, for both lifetimes.
What to do
- Put the parts that stay the same, like instructions, tools and reference material, at the start of the prompt. Caching reuses the unchanged beginning of a prompt.
- For instructions reused across longer gaps, set prompt_cache_options.ttl to 12h in the Responses API.
- Check your cache hit rate in OpenAI's Prompt Caching Dashboard.
Good to know
- A prompt prefix must be at least 1,024 tokens for GPT-5.6 and later to be cached.
- On GPT-6.1 Sol and GPT-6.1 Terra, cache reads cost 0.05x.
- OpenAI's guide: use 12 hours for stable instructions or reference material reused across longer gaps, and the default 30 minutes for conversations and agent steps that reuse context frequently.
- These are total rates, not fees on top of ordinary input.
Source
Checked on 2 Nov 2026. Prices, limits and settings change, so check the source before you rely on them.