OpenAI expands prompt caching for GPT‑6
OpenAI has extended prompt caching for its GPT‑6 models to a 30-minute reuse window and added a dashboard, miss diagnostics, explicit breakpoints, and prewarming — cutting latency and cost for long-running agents.
OpenAI says the GPT‑6 family shipped with a reworked prompt caching system that raises cache hit rates by default, with eligible shared prefixes now reused within a 30-minute window. Cached input tokens carry discounts of up to 90%.
The update is aimed at long-running agents that make chains of API requests carrying forward the same instructions, tool definitions, and prior context — codebase refactors, research documents, presentations.
Alongside the engine change, OpenAI added developer controls:
- A Prompt Caching Dashboard showing what share of an application's input is served from cache, hit rates over time, and an input composition chart comparing cached and uncached tokens.
- A diagnostics tool that compares a request against a recent response to explain misses. The returned
prompt_cache_diagnosticsobject names a reason (for example,tools_changed) and counts affected tokens. - Explicit cache breakpoints to pick which prompt prefixes get reused.
- Reasoning effort changes that don't break cache: on GPT‑6 models, developers can raise or lower effort between responses by appending a
configuration_updatewhile leaving request-level reasoning effort unchanged. - Prewarming, which processes known context — shared instructions, tool definitions, reference material — at startup rather than during the user's wait.
OpenAI also recommends keeping tool definitions, schemas, and ordering stable, using allowed_tools or tool_choice: none instead of removing definitions, and appending new developer messages late in the context to override older instructions.
The announcement does not give pricing changes, hit-rate benchmarks, or availability details beyond the linked guide and tools.