Better prompt caching for GPT-6

Persistent agents don’t make one call; they chain dozens, reusing instructions, tools, and context. OpenAI says GPT-6 ships an improved prompt caching system with higher hit rates by default, discounts on eligible shared prefixes reused within a 30-minute window, and new developer controls.

That sounds like plumbing. It isn’t. Caching is how long-horizon agents stay economically sane — and how latency stops eating the product.

We’re treating this as the quiet half of the GPT-6 story. Model scores get the headlines; cache hit rates decide whether agents can keep working after the demo.

OpenAI’s pitch is concrete: discounts of up to 90% on cached input-token reads, a Prompt Caching Dashboard to track hit rates and cached versus uncached composition, and a diagnostics tool to explain unexpected misses. Explicit cache breakpoints let developers choose which prefixes stay reusable. You can change reasoning effort between turns without busting the cache. Prewarming moves shared context off the user’s wait path.

Useful AI needs this kind of boring reliability. Agents that refactor codebases or draft long documents only work if the stack reuses what it already paid for. Customer quotes in the post — GitHub Copilot cutting fresh-processing share by more than 50% across billions of requests; Manus pushing hit rates from roughly 85% to consistently above 90%; another partner moving from 83% to 91% with explicit breakpoints — are the receipts that matter more than marketing adjectives.

The downside is concentration. When caching discounts deepen, switching costs rise. Apps get shaped around OpenAI’s prefix rules, tool-definition stability, and 30-minute eligibility window. That’s fine if you’re all-in; it’s a tax if you’re multi-homing. And diagnostics that surface miss causes also surface how fragile agent stacks remain — one tool reordering or instruction rewrite and the meter jumps.

I’m watching whether these defaults make multi-hour agents normal for mid-market products, not just labs with blank checks. If hit rates stay high in the wild, GPT-6’s price cuts land harder. If they don’t, the dashboard becomes a bill-shock monitor.

OpenAI product post on better prompt caching for the GPT-6 family, tied to agent workloads that carry forward shared instructions, tool definitions, and context across chained API requests.