On September 22, 2026, OpenAI announced on its official blog — in a post titled "Better prompt caching for GPT-6" — an improved prompt caching system for the GPT-6 family, along with new tools to help developers monitor and diagnose cache performance.
Here is what actually changed. GPT-6 models now get higher cache hit rates by default: eligible shared prefixes reused within a 30-minute window qualify for discounts of up to 90% on cached input tokens. OpenAI also launched a Prompt Caching Dashboard that tracks hit rates over time and uses an input-composition chart to compare cached and uncached tokens, helping teams spot drops in cache hits. When an unexpected cache miss occurs, the new diagnostics tool compares the request against a recent successful response to pinpoint whether the model, tools, settings, or input changed — and estimates how many tokens were affected.
The two changes agent developers will like most
First, developers can now set explicit cache-prefix breakpoints, choosing exactly which parts of a prompt get cached and where the cached segment starts and ends. Second, adjusting reasoning effort or adding and removing tools no longer automatically invalidates previously cached context. OpenAI's reasoning: GPT-6 is built to power persistent agents — from refactoring codebases to producing research documents — and these applications fire off sequences of API requests that all carry the same instructions, tool definitions, and context. Caching simply means that repeated content is only paid for once.
Who benefits most from the math
OpenAI cited GitHub data: over the past several months, caching improvements cut the share of prompt tokens needing fresh processing by more than 50% across billions of requests, while making Copilot respond faster. For teams running multi-turn conversations, document pipelines, or agent systems, input tokens are already the biggest cost line — a 90% discount on cached reads lowers the bill directly. But note the boundaries: the discount applies only to eligible shared prefixes within the 30-minute window; output tokens are excluded. If your app uses short prompts and single-turn Q&A, the caching dividend will be limited. Real savings still require optimizing prompt structure with the dashboard and diagnostics — turning it on is not enough by itself.
