Copilot’s usage API can now tell your own-key AI traffic from Copilot-hosted models

GitHub, owned by Microsoft, has updated the Copilot usage metrics API to report bring your own key (BYOK) activity separately from Copilot-hosted models, so admins can see who routes requests to their own providers or local models.

Two kinds of BYOK, two sets of fields

The API now distinguishes between local BYOK and enterprise custom models. Local BYOK is set up by developers in their own client: requests from Copilot CLI or the GitHub Copilot app go straight to a personal provider key or to a local model. GitHub lists Ollama, Foundry Local, LM Studio and llama-server as examples of tools that serve such models.

For local BYOK, user-level reports carry a flag, used_local_byok, showing whether a person made at least one such request that day. The totals_by_local_byok block gives session, request, prompt and token counts, split by byok_kind (local_managed, local_user, local_auto or remote_user) and by surface (copilot_cli or copilot_app). Aggregate reports add daily, weekly and monthly counts of active local BYOK users.

Enterprise custom models are configured by admins in GitHub settings, and their requests route through GitHub to the provider. They show up through a new boolean, is_custom_enterprise_model, on each model entry under totals_by_client. Because GitHub sees these requests on its servers, the field covers every client, IDEs included. A model used both as a hosted Copilot model and as a custom one appears as two entries, so the counts do not mix.

Both kinds appear in enterprise and organization reports, at user and aggregate level, in 1-day and 28-day reports. Access is open to enterprise owners, enterprise billing managers, organization owners and anyone whose custom role includes the View Copilot Metrics permission, provided the Copilot usage metrics policy is enabled.

Numbers that are a floor, not a total

GitHub is explicit about the limits. Local BYOK fields depend on client telemetry, so they undercount when telemetry is switched off, and token counts can fall short because some local providers do not report token usage. Local BYOK usage is not broken out by model name. Custom model counts, observed server-side, do not share this weakness. There is no backfill: days before the release lack is_custom_enterprise_model. Existing totals do not change, since BYOK traffic was already counted inside the CLI, app and client totals.

In practice, the old problem was that local and custom models often share names with Copilot models, so their usage blended into hosted totals. GitHub says the new fields let admins check BYOK activity against policy and judge cost per user more accurately.

Why the timing matters for budget owners

The release lands in a week when GitHub has been moving Copilot toward usage-based accounting. Under a model announced on October 28, enterprises working with a GitHub or Microsoft account team can get Copilot included for every GitHub Enterprise user, and AI usage is charged against an enterprise budget from the first request. GitHub says Copilot Standalone costs $10 per user per month and, like the bundled access, includes no monthly AI credits per seat. When spend is the main control, separating what Copilot serves from what a developer’s own provider bills becomes more than a reporting nicety.

The local side is also growing. A separate October 28 preview lets Copilot Fusion route between cloud models and on-device ones in Copilot CLI, the Copilot app and VS Code. On Windows, GitHub says, managed setup can use MAI Code 1.1 Flash through the Windows ML provider. The same post notes that routing decisions are still made in the cloud and that local models may have different safety protections than cloud ones. As more requests run on laptops, client telemetry becomes the only window onto them, which is exactly where the new local fields are weakest.

Our recent report on Copilot in JetBrains Air described cloud tasks billed to a Copilot plan, another sign that where a request runs and who pays for it now varies by tool.

What to watch in the first reports

The announcement gives no figures on how many developers use local BYOK, and it does not say how large the telemetry gap is. The first weeks of data will show whether admins find users bypassing policy, and whether the undercount is small enough to trust. The sharper test is a comparison: server-side custom model counts against client-side local counts for a similar population would show how much the lower bound misses. Until then, the local numbers are best read as evidence that BYOK use exists, not as a measure of how much of it there is. Documentation for the metrics API is linked from the changelog entry.