OpenAI Improves GPT-6 Prompt Caching With Up to 90% Savings and New Diagnostics

OpenAI improved prompt caching for GPT-6 models, adding higher default hit rates, new diagnostic tools, and finer developer controls aimed at long-running, persistent agents.

Details

• Higher hit rates: cache discounts now apply to eligible shared prefixes reused within a 30-minute window, delivering up to 90% savings on cached input tokens by default
• New diagnostics: a Prompt Caching Dashboard shows hit rates and token composition, alongside a diagnostics tool that compares requests against recent responses to explain why cache misses occur
• Explicit cache breakpoints: developers can specify exactly which parts of a prompt should be reused across requests
• Mid-conversation flexibility: reasoning effort can now be adjusted mid-conversation via configuration_update without invalidating the cache
• Tool management: keeping tool definitions stable while using allowed_tools to control availability preserves the cache rather than forcing a rebuild
• Cache prewarming: applications can prepare context during startup to cut user-facing latency

What happened next

The improvements target the same workloads GPT-6 Sol and Luna are built for — high-volume, cost-sensitive, and long-running agentic tasks — giving developers more precise control over caching behavior rather than relying solely on default heuristics.