Whoa, GPT-6's Prompt Caching Just Got a Huge Upgrade!

Hey everyone, it's Shii-chan! I just came across an update that made my day, so let's dive in together.

OpenAI News openai.com

What was announced?

OpenAI News published an announcement titled "Better prompt caching for GPT-6," describing a major overhaul of GPT-6's prompt caching system. Prompt caching is the mechanism that lets a shared prefix of your prompt get reused across multiple API calls, so you don't pay full price or wait full latency every single time.

According to the announcement, cached input tokens now get discounts of up to 90%, and eligible shared prefixes can be reused within a 30-minute window, a longer default caching window than before.

The story so far

Before this update, the caching window was narrower, so if you were running a long-lived agent that kept calling the API over an extended task, the cache could expire before you got to reuse it, which held back your hit rate. On top of that, when a cache miss happened, there wasn't a clear way to figure out why, so developers were mostly guessing.

What changes

With the wider 30-minute reuse window, longer-running agents and apps that repeat similar context across calls are much more likely to actually hit the cache. And every hit you land can save up to 90% on those cached input tokens, so the cost savings add up fast for anything that sends similar prompts repeatedly.

There's also a new Prompt Caching Dashboard that shows your hit rate over time and breaks down cached versus uncached tokens, so you can actually see whether your app is caching well instead of guessing.

Dive Deep

The announcement also introduces some concrete controls for keeping your cache alive:

• A prompt caching diagnostics tool that compares your request against recent responses to explain why a cache miss happened, down to the affected token count
• Explicit cache breakpoints, so you can choose exactly which part of your prompt prefix should be prioritized for reuse
• A configuration_update parameter that lets you change reasoning effort mid-conversation without breaking the cache
• Managing tools with allowed_tools or tool_choice instead of removing tool definitions, so your cached context survives even as an agent's available capabilities evolve
• Cache prewarming, which lets you prepare known context ahead of time so that work happens before the user is waiting, cutting latency

Wrap-up

• GPT-6's prompt caching now supports a 30-minute reuse window for eligible shared prefixes
• Cached input tokens can get discounts of up to 90%
• A new dashboard and diagnostics tool make hit rates and cache misses visible
• New controls like explicit breakpoints, configuration_update, and prewarming give developers more ways to keep the cache alive

If you're building long-running agents or anything that repeatedly calls the API with similar context, this is a genuinely useful update worth trying out!