September 23, 2026
Repeat GPT-6 requests will cost less: cache discount up to 90%
On September 22, OpenAI improved prompt caching in the GPT-6 API: cached input tokens receive discounts of up to 90%. GPT-6 routes repeated parts of requests to the cache itself, so `prompt_cache_key` is no longer needed. The cache activates at 1,024 tokens and retains the prefix for 30 minutes after it is written or reused.
OpenAI Developers
@openaidevs
We improved prompt caching in the GPT-6 API to make agents faster and less expensive. Now, more input tokens are cached by default and receive discounts of up to 90%.
· 77.8K views
Previously, developers specified `prompt_cache_key` to increase the likelihood that a shared prefix would be reused. In GPT-6, OpenAI routes requests to the cache itself. The `explicit` mode with a cache breakpoint remains for manually selecting the cacheable prefix.
Savings show up on the bill. One cache write and one full repeated request cost 1.35 times the regular input-token price instead of 2 without caching. Across 10 requests with full reuse, the cost is 2.15 instead of 10. With GPT-6 Sol, one million regular input tokens cost $2, while cached tokens cost $0.20.
In the GPT-6 API, caching is enabled by default from September 22.
