OpenAI overhauls prompt caching for GPT-6, adds dashboard and up to 90% discounts

OpenAI announced an upgraded prompt caching system for GPT-6, aimed at the multi-turn, long-running agent workloads the model family is built to handle. The update bundles a bigger discount, a wider reuse window, and new tooling for diagnosing why a cache misses.

What's new

OpenAI's own framing starts with the workload: "GPT-6 enables persistent agents to work for hours on complex tasks, from refactoring codebases to producing well-researched documents and presentations." Those agents make chains of API calls that repeat the same instructions, tool definitions, and context turn after turn, and OpenAI caches that shared prefix so it does not have to be reprocessed every time.

The discount on that reused, cached portion of the input now reaches "discounts of up to 90% on cached input tokens," reducing both cost and response time. OpenAI also extended the reuse window: eligible shared prefixes now qualify for the cache discount if they are reused within a 30-minute window, up from the shorter windows that made caching unreliable for slower-paced agent loops.

Alongside the pricing change, OpenAI shipped a Prompt Caching Dashboard for monitoring hit rates and a diagnostics tool that identifies why a given request missed the cache, something developers previously had to infer indirectly. The update also adds explicit cache breakpoints so developers can mark exactly where a prefix should be cached, the ability to adjust reasoning effort mid-conversation without invalidating the cache, and cache prewarming to cut latency on a request's first turn.

Context

Caching economics matter more for GPT-6 than for prior generations because OpenAI is explicitly positioning it for long-running, tool-using agents rather than single-shot chat completions. A codebase-refactoring agent or a research-and-write agent can burn through the same system prompt and tool schema across dozens of calls; without aggressive caching, that repetition becomes the dominant cost. OpenAI's developer changelog shows this caching update landing in the same stretch as the GPT-6 Sol and Luna model releases and the Agents API reaching public beta, all pieces of the same agent-infrastructure push.

Why it matters

A 90% discount on cached tokens changes the unit economics of running long agent sessions, which is the workload OpenAI is explicitly courting with GPT-6. The diagnostics tool addresses a real pain point: developers building agent pipelines have had no reliable way to see why a request fell out of cache, and unexplained cache misses translate directly into unexpected bills. Extending the reuse window to 30 minutes also better matches how actual agent loops behave, where a human or a tool might take a few minutes between turns rather than firing requests back to back.