GitHub announced on November 3, 2026 that the Copilot usage metrics API now reports bring your own key (BYOK) activity on its own. Until now, usage on a developer's own provider key or on a custom model often blended into the Copilot-hosted totals.
Admins of enterprises and organizations can now tell who is using which kind of model, and check that against policy and cost. The numbers show up in one-day and 28-day reports, per user and in aggregate.
Two Kinds of BYOK Are Now Reported
The new reporting covers two categories, because they behave differently. Developers configure local BYOK in their own tools, while admins configure enterprise custom models inside GitHub settings.
- Local BYOK: Copilot CLI or the GitHub Copilot app sends requests directly to the developer's own key or a locally hosted model. Models run through Ollama, Foundry Local, LM Studio or llama-server fall in this group.
- Enterprise custom models: traffic passes through GitHub on its way to the provider your admins picked. A flag on each model entry in the client totals marks them.
The New Report Fields
For local BYOK, a per-user flag says whether that person sent any local BYOK request on a given day. A new totals block, present in user and aggregate reports, counts sessions, requests, prompts and tokens by BYOK kind and by surface.
Kinds are local_managed, local_user, local_auto and remote_user, and surfaces are copilot_cli and copilot_app. Aggregate reports also list unique active local BYOK users per day, week and month.
For enterprise custom models, each model entry under the client totals gets a boolean named is_custom_enterprise_model. It reads true when the request went to a custom model set up by your enterprise or organization. GitHub observes this traffic on its servers, so IDEs and every other client are covered.
If one model name serves both roles, hosted by Copilot and as a custom model, it is listed twice so the counts stay apart. Custom entries carry actual request and token figures.
Why This Matters
Names of local and custom models frequently match those of models Copilot hosts, so older reports could not separate them. GitHub lists four benefits of the split:
- Governance: check who sends requests to personal keys, local models or custom models against your rules.
- Spend: separate what Copilot serves from what your own vendor bills or free local models handle.
- Visibility: one view of AI usage, whichever model or provider developers pick.
- Efficiency: compare people who rely on BYOK or custom models with the Copilot figures you already follow.
Limits to Keep in Mind
Existing totals stay the same, because this activity was already being counted. Local BYOK sits inside the CLI and app totals, and custom model traffic was already in the client totals. The new fields simply split it out.
Treat local BYOK figures as a minimum. They rely on client telemetry, so turning telemetry off hides activity, and token totals may fall short when a local provider omits them. Server-side custom model counts avoid this gap.
Other notes from GitHub:
- Custom model names go through the current name normalization, so names it cannot match are shown as unknown.
- Local BYOK usage is not broken out by model name.
- Earlier report days do not get the custom model flag retroactively.
- Enterprise owners, billing managers, organization owners and holders of a custom role with the View Copilot Metrics permission can read it, provided the usage metrics policy is on.
What It Means for Developers and Admins
If your team lets people connect their own keys or local models, you now have a way to measure that activity instead of guessing. A simple first step is to pull a 28-day report and look at which users and surfaces show local BYOK requests. Then compare that list with your internal rules on approved providers and data handling.
Remember the telemetry caveat before drawing firm conclusions: a low number may reflect disabled telemetry rather than low use. For cost reviews, treat enterprise custom model counts as the reliable figure and local BYOK as a floor.
Teams building automation on top of GitHub may also want to look at GitHub Agentic Workflows, now generally available, since agent usage is a growing part of the same spend picture.
Conclusion
The Copilot usage metrics API now tells BYOK and custom model usage apart from Copilot-hosted models. It is an additive change with no effect on existing totals. Admins who care about governance and cost should start reading the new fields in their next report.
Sources
See my articles more often when you search Google.
