OpenAI Improves Prompt Caching for GPT-6

OpenAI has published a technical note describing prompt caching improvements for GPT-6, covering higher cache hit rates, new diagnostics, explicit breakpoints and additional controls.

The underlying idea of prompt caching is straightforward. When repeated requests share the same prefix, keeping that prefix's state means later calls do not have to recompute it from scratch. The higher the hit rate, the more of a repeated prompt is reused and the more is saved.

Each of the four changes addresses a different part of that pipeline. Better hit rates mean more repeated prefixes are actually reused. Diagnostics make it possible to see whether a cache hit occurred, and if not, why not.

Explicit breakpoints hand the decision about where caching begins back to the developer. That boundary was previously inferred by the system, leaving teams to guess; declaring it makes the layering of long prompts, system instructions and fixed context far clearer.

The controls round out the policy layer, letting teams adjust caching behaviour by request type, business line or cost target instead of accepting a single default.

All of it points at two measurable outcomes: latency and spend. For teams running models in production, every gain in hit rate shows up directly in response times and invoices, especially with long system prompts or multi-turn conversations.

What to watch next is whether developers converge on shared conventions for explicit breakpoints, and how cache diagnostics get folded into existing observability stacks.