Usage

Request counts, token breakdowns, and latency for every inference request in a project.

arkor.ai/<org>/<project>/usage aggregates chat completions made through the project: endpoint traffic, playground messages, and CLI calls alike. One request equals one row, recorded when the request finishes. Only requests that pass ingress validation, authentication, and quota checks are counted — attempts rejected before dispatch (bad API key, malformed body, an Arkor 429 quota rejection) produce no row, so usage is not a record of every call your endpoint received. Requests that fail after dispatch do count: a request an external provider rejected with its own 429/529 backpressure, or one that failed mid-generation, appears as a failed row.

Summary tiles

• Requests in the selected range.
• Tokens, split three ways with no overlap: input, output, and reasoning. Output counts only visible completion tokens; reasoning tokens are separated out.
• Avg TTFT — time to first token, measured from provider dispatch, with the end-to-end figure (from when Arkor first received the request) shown alongside.
• Avg total — full generation time, same two measurements.

Daily chart

A per-day bar chart over the range, toggleable between request counts and stacked token breakdowns. Days are UTC buckets of request arrival time.

Breakdown table

Group requests by:

The endUser and session dimensions are high-cardinality and limited to 31-day windows; on the 90-day preset the dashboard falls back to the model breakdown and says so.

Filters

• Range presets: 24h, 7d, 30d, 90d (UTC).
• Endpoint filter: narrow every tile, chart, and breakdown to one endpoint.

Attributing traffic to your users

To make the endUser and session dimensions useful, send the attribution headers with your API calls. They are free-form labels, capped at 256 characters, and never affect authorization. See Attribution headers.