GPT-6 (version GPT-6 family) Turns Prompt Caching Into an Economics Problem Worth Solving

OpenAI's GPT-6 family supports persistent agents that can work for hours on complex tasks, but the quieter change is in how the platform prices and exposes shared context across requests. Cached input tokens can now receive discounts of up to 90 percent, and OpenAI has given developers new tools to see whether they're actually collecting that discount or paying full price without realizing it.

The discount only exists if the prefix survives

Prompt caching works by reusing a shared prefix of tokens across requests instead of reprocessing the same context every time. OpenAI caches shared context across API requests, and eligible shared prefixes receive the cache discount only when reused within a 30-minute window. For agents that run for hours, as GPT-6 is built to support, this creates a real constraint: the value of caching depends on how requests are sequenced and timed, not just on whether the content repeats. An agent that revisits the same context every few minutes benefits considerably. One that leaves longer gaps between similar calls resets the clock and pays full price again on the next request.

Seeing the hit rate instead of guessing at it

GPT-6's prompt caching provides higher default cache hit rates than prior setups, but the more consequential addition is the Prompt Caching Dashboard itself. It shows the share of application input served from cache and displays hit rates over time, which turns caching from an invisible backend behavior into something a team can monitor and be accountable for. Before this kind of visibility existed, a developer had no reliable way to confirm whether a given prompt structure was actually being cached or whether requests were silently missing. Visibility alone doesn't fix a broken cache, but it does tell you one is broken, which is the necessary first step before any diagnosis can happen.

What the diagnostics actually tell you about a miss

OpenAI's prompt caching diagnostics tool exists to help identify the causes of cache misses rather than leaving developers to guess. One documented example attributes a miss to changed tools, reporting 5629 comparison reusable tokens alongside 5629 cache-missed tokens for that case. The tool surfaces a likely category of change tied to the miss, which gives a developer a starting point for investigation rather than a black box output. That matters because prompt structures accumulate small changes over time, and without something pointing at a probable cause, teams are left reconstructing a change history by hand just to explain why a bill went up.

The controls that put you back in charge of reuse

Beyond diagnostics, GPT-6 gives developers direct levers over cache behavior. Explicit cache breakpoints let a developer choose exactly which prompt prefix should be treated as reusable, rather than relying entirely on automatic detection. Reasoning effort can be changed between responses without breaking cache reuse, which matters for workflows that adjust reasoning intensity by task but still want the surrounding context preserved in cache. And prewarming prepares known context before a request arrives, moving some of that work earlier in the pipeline. Stable tool definitions, schemas, and ordering also help preserve reuse, according to OpenAI, alongside these other levers. None of this is automatic; each control requires understanding your own request patterns closely enough to know which prefixes are worth protecting and which changes will break them.

Turning dashboard numbers into a workflow habit

The practical shift here is that prompt caching has moved from an opaque cost-saving side effect to a visible, tunable part of system design. A team building persistent agents on GPT-6 now has a dashboard to measure hit rates, a diagnostics tool to explain misses like the changed-tools case above, and explicit controls, including breakpoints, reasoning-effort flexibility, and prewarming, to shape reuse deliberately. Readers who want to practice this kind of hands-on diagnosis, reading a dashboard, tracing a miss back to its cause, and adjusting a prompt structure so the discount actually shows up on the bill, can work through it directly in NextLUCA's agent-building and applied AI training tracks, where sessions are built around real request logs rather than documentation alone.

Specs

• Cache discount: Up to 90% on cached input tokens
• Cache eligibility window: 30 minutes
• Comparison reusable tokens (diagnostic example): 5629
• Cache-missed tokens (diagnostic example): 5629

Sources

• OpenAI — OpenAI