<\/p>\n\n\n\n
OpenAI<\/a> has introduced improvements to Prompt Caching for its GPT-6 family, designed to help persistent agents reuse shared context across API requests and reduce response times and input processing costs.<\/p>\n\n\n\n The update targets applications where agents perform complex, multi-step tasks over extended periods. These include codebase refactoring, research-heavy documents and presentation creation.<\/p>\n\n\n\n Such applications often send multiple API requests that reuse the same instructions, tool definitions and context from earlier turns. OpenAI said its caching system stores eligible shared context so it can be reused across subsequent requests.<\/p>\n\n\n\n With GPT-6, OpenAI has introduced an updated caching system aimed at improving cache hit rates by default. Eligible shared prefixes can receive cache discounts when reused within a 30-minute window. Developers can receive discounts of up to 90% on cached input tokens.<\/p>\n\n\n\n OpenAI is also adding tools to help developers monitor cache performance, diagnose cache misses and control how much of a prompt is cached.<\/p>\n\n\n\n GitHub Chief Product Officer Mario Rodriguez said OpenAI’s caching system has helped GitHub Copilot reduce the share of prompt tokens requiring fresh processing by more than 50% across billions of requests, compared with its previous baseline. According to Rodriguez, this has contributed to a more efficient inference stack and faster time to first response for developers.<\/p>\n\n\n\n Developers can now use the Prompt Caching Dashboard<\/a> to monitor cache hit rates over time. The dashboard also provides an input composition view showing cached and uncached tokens.<\/p>\n\n\n\n Consequently, developers can identify changes that affect caching performance and investigate unexpected drops in cache hits.<\/p>\n\n\n\n For individual cache misses, OpenAI has introduced a prompt caching diagnostics tool<\/a>. The tool compares a request with a recent response and identifies changes involving models, tools, settings or input that may have prevented context reuse.<\/p>\n\n\n\n For example, the diagnostic output can identify a change in tool definitions and show the number of reusable tokens affected.<\/p>\n\n\n\n OpenAI is also giving developers more control over Prompt Caching through explicit cache breakpoints. The updated prompt caching guide<\/a> explains how developers can select prompt prefixes for reuse, as well as how changes to tools and inputs can affect cache eligibility.<\/p>\n\n\n\n In addition, GPT-6 models can now change reasoning effort between responses without breaking the cache. Developers can raise reasoning effort for more complex tasks or reduce it for routine follow-ups by appending a configuration_update, while keeping request-level reasoning settings unchanged.<\/p>\n\n\n\n OpenAI also recommends keeping tool definitions, schemas and their ordering stable when building agents. Developers can use allowed_tools to control which tools are callable, or set tool_choice to none when tools are not required, rather than removing tool definitions from the context.<\/p>\n\n\n\n Furthermore, new developer messages can be appended toward the end of the context to introduce updated instructions while preserving earlier reusable context.<\/p>\n\n\n\n