News · Source-backed update

Copilot usage metrics: agent activity is not your bill

GitHub adds client and custom-tool breakdowns to Copilot usage reports. Check attribution, permissions and missing telemetry before interpreting AI agent activity.

Edited by Clinton FeyisitanPublished Provider event / observation: 2026-10-23

Review scope: Selected official pages read 26 October. Documentation review only; no account, API, import, payment or send tested.

Which AI agent used the allowance?

GitHub’s 23 October update adds client breakdowns to Copilot usage reports, including activity from third-party clients such as OpenCode. The scope is model traffic routed through Copilot, not every action in every coding tool. One-day and 28-day reports include requests, sessions, tokens and model breakdowns at enterprise, organization and user levels.

Requests count model calls, not necessarily user prompts. Client attribution is best-effort integration identification. Access still requires the relevant metrics permission and enabled policy.

Named skills and MCP tools have coverage limits

The 22 October update adds named skills, custom agents, MCP tools, slash commands and plugins, with up to 25 named entries before an “other” group. CLI-based reporting needs version 1.0.97 or later; SDK reporting needs 1.0.21. Native IDE chat does not report those names. Historical activity is not backfilled, and missing telemetry can undercount activity.

A worksheet for interpreting the report

This is a proposed review worksheet, not a report of an executed account test. Use aggregate information and keep private tool names or user-level records out of public reports.

  1. Fix the scope: record organization, complete dates, client versions and enabled reporting policy.
  2. Separate units: put model calls, sessions and tokens in different columns. Do not label any of them completed tasks.
  3. Mark missing coverage: note clients and periods that cannot report. A blank field is not proof of zero use.
  4. Reconcile separately: compare the billing account’s actual invoice and allowance definition. Usage telemetry alone does not establish an amount charged.
  5. Check outcomes: record reviewed code and accepted work independently. More agent activity does not demonstrate productivity.

Continue with the AI coding bill worksheet or the dated Copilot profile. No plan recommendation or score changes as a result of this announcement.

Sources and scope

Provider statements are attributed above. Proposed checklists are not executed test results. Historical receipts, recommendations and scores remain unchanged.