OpenAI has rolled out an upgrade to prompt caching across its GPT-6 model family, introducing higher cache hit rates and new developer controls designed to make persistent AI agents faster and less expensive to operate. The company detailed the changes in an official announcement.

Prompt caching allows applications to reuse computation whenever multiple API requests share the same instructions, tools, or context. Under the updated system, cached input tokens can qualify for discounts of up to 90%, and GPT-5.6 and later models can reuse eligible cached prefixes for at least 30 minutes after their most recent use. For agent-style applications, which tend to resend the same instructions, tool definitions, and conversation history with every turn, how much of each request can be served from cache has a direct bearing on both operating cost and response time.

The improvements are aimed in particular at agents that work over extended periods and repeatedly carry large volumes of context between requests, such as coding agents refactoring a codebase or systems producing long research documents — workloads where the same context may otherwise be reprocessed many times over the course of a single task.

Alongside the model-side changes, OpenAI has launched a Prompt Caching Dashboard that lets developers track cache hit rates and compare cached and uncached token usage. A separate diagnostics tool can identify changes to models, tools, settings, or inputs that caused a request to miss the cache. Together, these tools give teams a way to measure how much of their real-world traffic is actually cacheable, turning cache behavior from an invisible infrastructure detail into something that can be monitored and tuned.

Developers can also now set explicit cache breakpoints to control which portions of a prompt are reused, making it possible to keep stable content such as system instructions and tool definitions cached while more volatile segments of a prompt change. GPT-6 models further allow reasoning effort to change between responses without invalidating previously cached context when configured appropriately.

Another addition is cache prewarming, which lets applications process shared instructions, tool definitions, or reference material before a user sends a request, reducing the amount of processing required before the model begins responding.

The scale of the opportunity is already visible in customer data: GitHub said the improvements to OpenAI's caching infrastructure reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models — a sign of how large the cacheable share of real-world workloads can be.

The update builds on the prompt caching feature OpenAI introduced in 2024, which initially offered automatic discounts for repeatedly used prompt prefixes. The GPT-6 system extends those capabilities with longer reuse windows and more direct developer control over cache behavior. As agent sessions stretch from single exchanges into hours-long tasks, the share of a workload that can be served from cache is becoming a practical factor in how teams structure prompts and manage API costs.