OpenAI upgrades GPT-6 prompt caching with up to 90% token discount

OpenAI has upgraded the prompt-caching mechanism in the GPT-6 API, promising faster agent performance and lower costs through higher default cache-hit rates. According to the announcement, more input tokens will qualify for the cached-input discount of up to 90%, translating into direct savings on every call that reuses context.

New dashboard and diagnostics API

Alongside the automatic improvement, the company is launching a dedicated dashboard at platform.openai.com that shows in real time which parts of a prompt are being reused and which are preventing a cache hit. A new diagnostics API endpoint lets developers query what prompt changes blocked reuse and returns an estimate of the affected token count — a tool designed to shorten the trial-and-error loop when the cache behaves unexpectedly.

Granular cache-behavior controls

Developers now receive three explicit tuning knobs: cache breakpoints that define which prompt prefixes get cached; the ability to adjust reasoning effort and tool availability without losing the stored context; and a prewarm mechanism that loads shared context ahead of time so the first response returns faster. Each capability is intended to enable finer adaptation to multi-turn application architectures and agents that maintain state over long sessions.

What this means for developers

In practice, the change shifts the optimization burden from application code to OpenAI's infrastructure. Instead of building custom caching logic or manually trimming prompts, teams can rely on higher default hit rates and use the dashboard and API to spot bottlenecks when they appear. The 90% discount on cached tokens remains a theoretical ceiling; actual realization depends on each application's specific usage patterns, and the new tools are designed precisely to measure that gap.