OpenAI's GPT-6 prompt caching adds breakpoints and diagnostics

Alongside its new models, OpenAI reworked prompt caching for GPT-6, the mechanism that reuses a repeated prompt prefix instead of recomputing it. The change is aimed at developers who had been reverse-engineering the behavior to save money. OpenAI says cache discounts now apply to eligible shared prefixes reused within a 30-minute window, and that hit rates rise by default: one customer went from about 85 percent to consistently above 90 percent in under a week, and another moved from 83 to 91 percent after switching session agents to explicit breakpoints.

The more useful part is control. Explicit cache breakpoints let you pick which prefixes to reuse through prompt_cache_options.mode and prompt_cache_breakpoint on GPT-5.6 and later, so a long system prompt or tool definition can be cached deliberately rather than by luck. A new diagnostics tool explains why a request reused fewer tokens than expected, pointing to the change in model, tools, settings, or input that broke reuse, and a dashboard tracks how much input is cached over time. The pricing stays familiar: cached input tokens cost 10 percent of the uncached rate, while cache writes are billed at 1.25 times it.

Why it matters

For anyone running long, repetitive prompts through GPT-6, this turns caching from a black box into something you can steer and audit. If your agent reuses a large system prompt on every call, placing an explicit breakpoint and watching the diagnostics can cut real cost and latency, so it is worth a pass through your hottest endpoints.