Spend and usage by your own customers (TRA-1307).

Send your customer's organization id as the X-Tenant-Id header on each inference request, and its display name as X-Tenant-Name. This returns one row per X-Tenant-Id your requests carried between start and end. Rows are priced at the team's rates and sorted by tokens, most first. tenant_name is the name that tenant sent most recently. Requests sent without the header form one row with a null tenant_id, so the rows add up to what /v1/billing/spend reports for the same window and model. A dedicated GPU day is split across the tenants that ran on the GPU by their share of each 15-minute cell, so tenants that shared the GPU share its cost rather than each paying for it.

The window and model work as they do on /v1/billing/spend. A 503 means the gateway has not finished building per-tenant usage for the window yet. Its detail names the first day it can answer for.