Skip to main content
2026/8/21 11:39 (UTC+8) · Service Notice · OpenAI / Anthropic 📊 Console logs now break out cache tokens — hits and writes are visible per call The log list used to show a single prompt-token total, and cache hits were only visible after opening the Cache Billing Details panel. That column now splits into three numbers: uncached input tokens on top, a green down arrow for cache hits (cache read), and an orange up arrow for cache writes. Hovering any of them shows the exact value, for example “命中（读取缓存）：32,815 tokens” (Hit / cache read: 32,815 tokens).
Console log list showing claude-sonnet-4-6 rows where the prompt token column displays uncached input, a green down arrow 33.4K, and an orange up arrow 542

The new cache token breakdown in the log list — green ↓ is cache read, orange ↑ is cache write

Log row tooltip reading 命中（读取缓存）：32,815 tokens

Hover shows the exact value: Hit (cache read): 32,815 tokens

For models billed with cache tiers such as OpenAI and Claude, this makes billing verifiable row by row: the cache-read portion is charged at the cache rate and the cache-write portion at the write rate, and both reconcile directly against the actual charge on the right. Previously, confirming a hit meant opening Cache Billing Details — the usage cache fields echoed back in the API response are not always accurate — and it can now be read straight off the list. How caching is enabled and what makes a hit are unchanged; see Does APIYI Support Cache Billing?.
← Back to Live Updates · 📚 Monthly Archive