GPT‑6 introduces an improved prompt caching system that delivers higher cache hit rates by default and offers discounts of up to 90% on cached input tokens. The system reuses shared prefixes within a 30‑minute window, reducing latency and costs for long‑running agent tasks. New tools include the Prompt Caching Dashboard to monitor hit rates and a diagnostics tool to analyze cache misses by comparing requests and identifying changes in models, tools, or inputs.

Developers can now use explicit cache breakpoints to choose which prompt prefixes to cache and adjust reasoning effort without breaking the cache. To preserve cache stability, keep tool definitions and schemas consistent, use allowed_tools for relevance, and append new instructions via developer messages rather than overwriting old ones. Prewarming prepares known context ahead of time, moving processing out of user wait time.

These optional controls build on default performance, letting developers tailor caching to their workload. Start by monitoring hit rates, investigating misses, and following the updated prompt caching guide or using Codex to review code.