- GPT-6’s prompt caching now delivers higher cache hit rates plus new diagnostics and controls.
- Explicit cache breakpoints let developers pin what gets reused across calls.
- The target workload is persistent agents that run for hours on complex tasks.
- Caching is now a first-order lever on agent economics, alongside raw token price.
What Happened
OpenAI published an update on September 22, 2026 detailing better prompt caching for GPT-6: higher cache hit rates, new diagnostics, explicit breakpoints, and controls that reduce latency and costs, per the company’s post. OpenAI frames the work around GPT-6’s ability to run “persistent agents [that] work for hours on complex tasks, from refactoring codebases to producing well-researched documents.”
Why It Matters
For agentic workloads, the same system prompt, tool definitions, and context get resent on every step — so cache efficiency, not list price, increasingly decides what an agent actually costs to run. Cached tokens are billed at a fraction of fresh input across the industry, which is why caching improvements compound: an agent looping for hours re-reads far more than it writes. Model-by-model economics are tracked in our LLM Price Index.
Technical Details
The additions are practical: diagnostics that show why cache hits are missed, explicit breakpoints so developers control cache boundaries instead of relying on automatic prefix detection, and higher hit rates from the caching layer itself. Explicit breakpoints matter most for agent frameworks, whose dynamic context tends to defeat automatic prefix caching.
Who’s Affected
Developers running long-lived agents get lower bills and latency without changing models. Agent-framework maintainers gain the control primitive they have been working around. Competing providers, who market their own caching discounts, face pressure to match the diagnostics and explicit controls.
What’s Next
Real-world numbers from agent-heavy teams will show how much the higher hit rates translate into cost reduction. Expect cache behavior to join context length and token price as a standard line item in model comparisons.
