What changed
OpenAI announced expanded GPT-6 prompt-caching controls and diagnostics on September 22. Repeated context deserves its own performance budget.
Why it matters
Agent requests often repeat instructions, tool definitions and repository context. Caching can change the economics of that repetition, but a hit rate alone misses whether cached material is still relevant. Separate stable instructions from frequently changing task state. Never assume cached context is a durable memory or a guarantee that stale information is corrected.
What to check next
Instrument requests before changing prompt layout. Record reusable prefix length, observed cache hits, latency and completed-task cost. Change one variable at a time and include a cold-cache run. Treat any provider saving as a measurement to reproduce, not a forecast.
Event dates refer to the original announcement. Publication dates show when this article first appeared on Codorc.
