GPT-6 gains cache monitoring, miss diagnostics, explicit breakpoints and prewarming, with discounts of up to 90% for eligible cached input tokens.
OpenAI has upgraded prompt caching across the GPT-6 family, targeting persistent agents that make sequential API calls while repeatedly sending the same instructions, tool definitions and prior context. The system reuses computation for shared prompt material, which OpenAI says can reduce latency and lower the cost of eligible cached input tokens.
The company advertises discounts of up to 90% on cached input tokens, but the benefit is conditional rather than a general reduction for all input. It applies to eligible shared prefixes reused within a 30-minute window. OpenAI also says GPT-6 applications should receive higher cache-hit rates by default under the updated system.
A new Prompt Caching Dashboard shows how much of an application's input is served from cache. Developers can monitor hit rates over time and compare cached and uncached token volumes. A separate diagnostic tool compares a request with a recent response, identifies changes that blocked reuse and estimates the number of affected tokens.
OpenAI has also introduced explicit cache breakpoints, letting applications designate prompt prefixes for reuse. GPT-6 can change reasoning effort between responses without invalidating cached context when the developer appends a configuration update and keeps the request-level reasoning-effort setting unchanged.
Cache prewarming provides another optional control. It processes instructions, tools or reference material before a user's request, with the aim of reducing perceived response latency. This feature supplements automatic caching and requires the relevant context to be known in advance.
Cache reuse remains sensitive to request structure. Changes to the model, tools, settings or input can cause misses, while modifying or reordering tool definitions can make previous context non-reusable. OpenAI recommends retaining tool definitions, schemas and ordering; limiting callable tools through allowed_tools or tool_choice; and placing revised instructions in new developer messages.
The update is therefore most relevant to agents with long, stable context and frequent consecutive calls. Applications that can separate durable instructions and schemas from changing user data have more opportunity to reuse computation. Benefits will be narrower when prompts are continually restructured or calls fall outside the 30-minute window, and OpenAI presents its customer results as organization-specific rather than universal guarantees.
Sources
Originally published by AI Feed: OpenAI upgrades prompt caching for GPT-6 agents.
Top comments (0)