OpenAI improves prompt caching for GPT-6 with diagnostics and breakpoints

OpenAI prompt caching GPT-6 is receiving changes aimed primarily at developers of AI applications and agents. On September 22, 2026, the company announced that shared input prefixes can qualify for a cache discount when reused within a 30-minute window. The updates also include a cache usage dashboard, cache-miss diagnostics, and explicit cache breakpoints for selected models.

Prompt caching makes it possible to reuse an already processed, stable portion of an input instead of processing it repeatedly with every API call. With longer prompts, this could include system instructions, tool definitions, or extensive context that does not change between agent steps.

OpenAI prompt caching GPT-6 and shared prefixes

According to OpenAI, shared input prefixes qualify for a cache discount if they are reused within 30 minutes. A prefix means the initial, identical portion of an input. For developers, how they arrange a prompt is therefore important: stable instructions and context should come before frequently changing data.

The change applies to the API layer. It is not a new feature or capability intended directly for regular ChatGPT users. Its significance mainly concerns operators of applications that repeatedly send a large amount of the same context to the model.

Dashboard and cache-miss diagnostics

OpenAI added a cache usage dashboard and cache-miss diagnostics. When the cache is not used, the diagnostics may indicate the reason for the miss as well as an estimate of the number of affected tokens. Developers should therefore gain a better overview of whether stable parts of their inputs are actually being reused.

Such data may help reveal changes in prompt structure or request configuration that prevent cache reuse. The documentation also describes diagnostic types for cache hits and cache misses, along with the relevant token counters.

Explicit breakpoints from GPT-5.6

For GPT-5.6 and newer models, OpenAI documents explicit cache breakpoints. Developers can use them to mark the stable part of the context they want to cache without having to cache the frequently changing end of the prompt.

This option is intended for cases where the beginning of a long request remains the same, while the final section contains current data, new user input, or the result of the agent’s previous step. The breakpoint is intended to allow these two parts to be separated more precisely.

Cache write and read pricing

The documentation states that writing to the cache costs 1.25 times the standard price for input tokens. Subsequent reads from the cache are charged at 0.1 times the standard price for input tokens. Cache is therefore not an automatic saving on every call: the benefit comes from repeatedly using the same prefix.

For multi-step AI agents, these mechanisms may be significant when the same instructions, tool definitions, and context are sent repeatedly. However, OpenAI has not published an independently verified, generally applicable figure for an increase in cache hit rate or savings for all customers.

In connection with caching, the company also cites customer claims, including those from GitHub Copilot and Manus, about reduced costs or a higher share of cache hits. These are customer claims presented by OpenAI, not universal results.

What to watch next

• pricing terms and availability of the updates for specific GPT-6 variants,
• independent measurements of savings, latency, and cache-hit rates in real-world deployments,
• the practical limits of the cache’s 30-minute lifetime and the effect of changes to the model, tools, or service tier on cache reuse.

Verified and updated: September 23, 2026 06:21