Cache reporting
Cache-read tokens, report status and statistics for the selected time window.
On this page
Per-request fields
Upstreams may return cached_tokens and cache_write_tokens in prompt_tokens_details; fields differ by protocol and model. Missing fields mean unreported, not zero hits. The console shows cache_read_tokens and cache_read_status: reported, not_reported, unknown, invalid or not_applicable.
{"prompt_tokens_details":{"cached_tokens":60,"cache_write_tokens":20}}Window statistics
Statistics cover all filtered chat rows, not just the current page; dates use UTC. Reported-sample ratio = valid reported cache reads / those samples' input tokens. Coverage = reported input / all-window input. Observed share = observed cache reads / all-window input. Missing reports are separate, so observed share is not a true overall hit rate; no valid reports shows —.
Usage tips
Stable system prompts, document prefixes and tool definitions may help caching, depending on upstream policy. Discounts depend on actual settlement; the platform promises no fixed hit rate.