Better Prompt Caching for GPT-6

For developers on the OpenAI API: GPT-6 models reuse repeated prompt content more often, with new tools to see and fix what isn't being reused.

OpenAI stores prompt content that repeats across requests and charges up to 90% less for those cached input tokens
GPT-6 models get a new caching system with higher hit rates by default, and repeated prompt openings now qualify for the discount if reused within 30 minutes
A diagnostics tool explains why a request missed the cache (for example, the tool list changed) and estimates how many tokens that cost you
Explicit breakpoints let you choose which part of a prompt to cache, and prewarming loads known context before the first user request arrives
On GPT-6 you can raise or lower reasoning effort between turns without losing the cache
The Prompt Caching Dashboard tracks hit rates over time alongside these new controls

Long-running agents that resend the same instructions, tool lists, and context on every request.

Your GPT-6 API bill or response times are climbing as agent sessions get longer.

Developers and platform teams running agents on the OpenAI API at volume.

Join the newsletter to get ahead with AI.

I share strategies, new modules, free guides, and tools to get you better at using AI in your work.