September 23, 2026
Repeat GPT-6 requests will get cheaper: cache lasts 30 minutes
On September 22, OpenAI increased GPT-6's cache hit rate and kept 30 minutes for reusing shared prefixes. Cache reads cost 0,1 of the input-token price.
Previously, OpenAI enabled prompt caching automatically if the beginning of a request contained at least 1 024 visible input tokens. Now GPT-6 recognizes a shared prefix more often, so the repeated part of a long request is not processed again.
Control points. For explicit caching, enable `prompt_cache_options.mode: "explicit"` and add `prompt_cache_breakpoint` to a content block. One cache write and one full read cost 1,35 times the standard input price, versus 2 without caching. After the last point, the changing tail of the request is charged at the standard rate.
In the Responses API, pass the base response through `comparison_response_id`, then check `prompt_cache_diagnostics` and the number of cached tokens. A request with `prewarm: true` warms up the shared context before the main request without generating a response.
Early figures. The Manus team raised its hit rate from around 85% to a steady 90% in less than a week. At Wordsmith, switching to explicit points raised it from 83% to 91%, cut cache writes by around two-thirds, and reduced inference costs by 36%.
The new Prompt Caching Dashboard already shows the share of input tokens from cache, hit rate over time, and the ratio of cached to regular tokens.
Source
