OpenAI Lifts Default Prompt-Cache Hit Rates for GPT-6

Prompt caching lets an AI provider skip reprocessing input it has seen recently and charge less for those tokens, which suits agents because they resend the same instructions, tool definitions and conversation history on every step. OpenAI's developer account says it has improved prompt caching in the API for GPT-6 so that more requests hit the cache by default, which means more input tokens get the cached-input discount of up to 90% and agents run faster and cost less (OpenAIDevs on X).

By default is the useful part. In an earlier OpenAI Build Hour session on prompt caching, which discusses GPT-5-era models, an OpenAI solutions engineer explained that a cache hit requires a request to begin with exactly the same input, in the same order, as an earlier one. She also said that when thousands of requests share the same opening they are spread across machines, so hit rates could drop unless developers passed an optional key grouping related requests onto the same machine. A higher default hit rate suggests fewer of the savings depend on that kind of tuning. The first post of OpenAI's thread, the only part collected here, does not say what changed for GPT-6 or how much hit rates rose, so how much any particular workload saves is unknown from this source. The post links to a video that could not be transcribed, and OpenAI's blog post on the change was available here only as a one-line description mentioning new diagnostics, explicit breakpoints and other cost and latency controls.