GPT-6 gets improved prompt caching — diagnostics, explicit breakpoints, and a lower price

Prompt caching for GPT-6 reworked: explicit breakpoints and higher cache hit rate ⚡

In agentic scenarios, it's not single requests that count, but repeated ones. It's precisely for those that OpenAI reworked prompt caching for GPT-6 — on September 23, 2026, the page "Better prompt caching for GPT-6" announced an increase in cache hit rate, that is, the share of cache hits.

What exactly was rolled out:

— cache hit rate increased;

— new cache diagnostic tools;

— explicit breakpoints, explicit cache break points;

— controls that cut latency and cost.

The effect was stated without fluff: repeated requests become cheaper and faster. And it's repetitions that any agentic pipeline relies on.

What is most often overlooked: diagnostics and explicit breakpoints turn the cache from a black box into a manageable one.🔍 Before this, there was no way to catch where the cache break points were. Now they are visible and can be set manually.

A breakdown of caching for GPT-6