What's this about?
OpenAI now exposes cache hit dashboards and miss reasons, lets developers set prefix breakpoints and supports prewarming recurring instructions and tools.
Why does prompt order matter?
Caches reuse an identical beginning of a prompt. Stable tool schemas and append-only instruction changes preserve that prefix; deleting or reordering them can force expensive reprocessing.
What's the practical takeaway?
Official product documentation; discounts reach 90% on cached input.
- Track hit rate rather than assuming caching works.
- Use `allowed_tools` instead of rewriting tool definitions every turn.
- Change reasoning effort through configuration updates without invalidating context.