Usage tells a story.

The service counts retained successful inference requests created inside the selected half-open UTC window, using external-purpose credentials and confirmed metering. The current published, enabled text-model set defines eligibility. Failed requests and internal credentials are excluded; this is sandbox metering, not global market traffic.

Each applied model, public provider group, task kind, window and metric is sent to the service for recomputation. The complete matching aggregate is ranked before the top-100 limit; equal values use model-ID code-unit order. Totals and trends include all matching retained records. A provider group reflects current reviewed public model associations, not historical supply routing.

Requests are logical successful requests; USD values are recorded sales amounts, not verified revenue or profit. Any missing amount makes the affected group, total or interval unknown. An empty bucket means no observed retained record, not proven zero traffic. Retention completeness is unknown; image, video, audio, BYOK and PDF-specific task dimensions are outside this dataset.