OpenAI ships cache diagnostics: see why your API calls miss cache
A silent cache miss means paying full input price on every call. The Responses API now names which of nine reasons broke it — plus two Vercel cost levers.
Copy markdown
Nine reasons, one field
Prompt Cache Diagnostics went GA in the Responses API for GPT-5.6 and newer. When a call misses cache, it now returns the first classified reason out of nine — a model, cache-key, service-tier or tool change, a text-format, reasoning-effort or verbosity shift, context compaction, or an altered prefix. It adds no extra cost or rate-limit charge, so you stop guessing why a cached prompt suddenly bills at full rate.
A hint, not a receipt
The classifier compares your request against a recent baseline response, so a 'cache_hit' label doesn't prove every token was reused — keep reading usage.input_tokens_details.cached_tokens for the real count. And it's Responses-API-only: no Chat Completions, no GPT-5.5.
Elsewhere: Vercel trims two cost levers
Vercel Sandbox routing lookups now serve from region-local data, up to 18x faster globally — noticeable if your agents cold-start sandboxes far from home. And Flat Rate CDN reached GA for Pro teams: a predictable monthly bill with built-in spike protection instead of usage surprises.