OpenAI Releases Prompt Caching Optimization Guide and Diagnostic Tools for GPT-6 Models

OpenAI has released new features and guides to optimize Prompt Caching performance for GPT-6 models. Key features include Explicit Cache Breakpoints, which allow developers to explicitly select reusable prompt prefixes; Prewarming, which reduces latency by pre-processing shared instructions and tool definitions at application startup; and the ability to adjust reasoning effort across GPT-6 models without breaking the cache during responses.

Additionally, developers can diagnose and resolve caching issues using the Prompt Caching Dashboard, which tracks cache serving rates, and the Diagnostics Tool, which analyzes the causes of unexpected cache misses.

Customer Use Cases

• Codex: Leveraged OpenAI's diagnostic tools and dashboard to improve cache hit rates by a few percentage points (p.p) and reduce costs by 20%. CTO Arian Hanifi stated that this made it economically feasible to create conversation forks for background tasks while reusing shared context.
• Manus: Collaborated with the OpenAI engineering team to improve cache breakpoint placement, raising cache hit rates from approximately 85% to over 90% in less than a week. Team Leader Bin Fan reported that this further reduced inference costs in production environments.
• Strawberry Browser: Switched session agents to explicit cache breakpoints, increasing benchmark cache hit rates from 83% to 91% in less than a week. AI Engineer Eugene Mikhantyev reported that cache write volume decreased by approximately 2/3 and inference costs were reduced by 36%.