GPT-6 Prompt Caching Boosts Hit Rates, Adds Diagnostics

OpenAI has announced an improved prompt caching system for its GPT-6 family, designed to help persistent agents that make long series of API requests run faster and more cheaply. The system caches shared context—instructions, tool definitions, and earlier conversation turns—so that repeated computation can be reused. Developers can receive discounts of up to 90% on cached input tokens, and eligible shared prefixes now get cache discounts within a 30-minute window.

The update includes new operational tools. A Prompt Caching Dashboard shows how much input is served from cache, tracks hit rates over time, and compares cached and uncached tokens. A diagnostics tool lets developers compare a request with a recent response to identify why a cache miss occurred—for example, changes to tools, settings, or input—and estimates the number of affected tokens.

OpenAI also introduced controls to help developers optimize caching. Explicit cache breakpoints allow developers to choose which prompt prefixes to reuse, and GPT-6 models now let developers adjust reasoning effort between responses without breaking cache. The company recommends keeping tool definitions stable and using developer messages to append new instructions, and it suggests prewarming known context during startup to move processing out of the user's wait time.

Partner quotes in the announcement describe real-world gains. GitHub Copilot says it reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests. Manus reported raising its cache hit rate from roughly 85% to above 90% in under a week, and Wordsmith said cache hit rates on its evaluations rose from 83% to 91%, cutting inference costs by 36%. These are vendor-reported figures from a single source, so they should be treated as such.