GitHub announced on October 23 totals_by_client in the Copilot usage metrics API. It expands visibility into tools using Copilot as their model provider, including OpenCode.

The breakdown covers successful calls, sessions and tokens by client, with model details. Calls are not prompts: one request can invoke a model several times. Client identification relies on the integration ID supplied, a best-effort grouping rather than verified identity.

Looking beyond the editor

MnzAI Labs' assessment is that this enables a more precise management question: which tool consumes available AI capacity, and for what work?

Consider a team switching between an editor and a terminal tool. Editor activity alone can give an incomplete picture. This is an illustrative measurement scenario, not a newsroom product test.

A useful assessment separates three questions:

Question

Where does consumption occur?

Evidence to seek
Distribution of calls and tokens
Question

Was the work completed?

Evidence to seek
Accepted deliverables and rework
Question

Was the result worthwhile?

Evidence to seek
Total time, quality and observed cost

More tokens do not establish better output. Higher consumption might accompany harder tasks, repeated attempts or a changed workflow. Interpretation requires that additional context.

Bringing the data into reporting

The API documentation describes daily and 28-day reports accessed through temporary download links. Access requires appropriate permissions and an enabled metrics policy.

For existing dashboards, our recommendation is to check ingestion compatibility with the new field, preserve each measure's origin and record collection dates. An increase following broader coverage deserves investigation before being labeled genuine adoption growth.

Periods must also align. GitHub's reconciliation guidance says differences in windows, scope and processing can explain discrepancies between dashboards and exports. Recent telemetry can take three full UTC days to finalize.

Better observability supports better questions. Productivity conclusions require evidence of delivered work.

Engineering leaders can use the measurement to guide support and experiments across comparable tasks, without turning call volume into a ranking of individuals.

Official sources in English. Announcement: October 23; editorial review: October 24, 2026. The newsroom did not access customer usage data.