Get Token Report

Where the LLM spend goes, and whether the stack is healthy (admin).

Tokens per day stacked by surface and service, the heaviest tenants, the model mix, the cached-token share, and the share of calls whose outcome is not ok.

non_ok_share is the health signal, not the token total: outcome='unknown' names a degraded remote leg — a call that never came back spent almost nothing and broke something, so it is invisible in spend and visible here.

Windowed, never all-time, and window_start is in the response so the totals are never read as more than they are: llm_usage_event has no prune job, so an all-time aggregate is a full scan of a table that only grows. Capped at 90 days for the same reason.

Charged tokens exclude cached ones and rows carrying an uncharged_reason — the semantics repos/llm_usage already settled, so this page's totals agree with the billing the same platform does. Cached tokens are reported as their own number rather than netted out: they are the discount worth growing.