If you've ever tried to answer "which tool is spending our Copilot budget?" from GitHub's own data, you know the answer used to be "the IDE, probably". Developers who used Copilot as the model provider behind Claude Code or OpenCode mostly didn't show up by tool at all.
October changed that. It also changed three other things about the usage metrics, and one of them means some of your September charts were wrong. Here's what moved, and the one distinction that keeps all of it from being misread: metrics and billing are separate ledgers.
What changed in October
GitHub shipped four changes to Copilot usage metrics between October 6 and October 23, 2026. Read together they're a big step for visibility, but each one has a scope note worth knowing before you build a report on it.
| Date | Change | Where it shows up | Source of the data |
|---|---|---|---|
| Oct 6 | Agent activity undercount in SDK-based IDEs identified, fix rolling out | Dashboard and API | Client telemetry |
| Oct 22 | Skills, custom agents, MCP servers, slash commands and plugins reported by name | API reports | Client telemetry |
| Oct 22 | Industry comparison section in the Copilot impact dashboard | Dashboard only | Opt-in peer aggregates |
| Oct 23 | totals_by_client: requests and tokens per tool and per model |
API reports | Server-side |
The last column is the one to keep in your head. Client telemetry can be blocked, outdated or missing. Server-side counts can't see inside the editor, but they don't depend on it either.
totals_by_client: tokens by tool, for the first time
On October 23, GitHub added a totals_by_client array to the enterprise, organization, enterprise-user and organization-user reports, in both 1-day and 28-day versions (GitHub Changelog, October 2026). Each entry is one client: opencode, claude_code, codex, zed, copilot_cli, github_app, cloud_agent, code_review, vscode, jetbrains, or other.
Every entry carries a request_count, a session_count and token_usage with prompt tokens, output tokens and average tokens per request. Under that sits totals_by_model, the same metrics broken down by model, with a model_selection_mode telling you whether a model was user_selected, picked by auto, or product_selected.
Why does that matter? Because GitHub counts these requests on the server. Tools like OpenCode often don't send editor telemetry, so their usage was missing from totals_by_ide entirely. Now it isn't.
Three caveats from the changelog itself, and each one changes how you'd use the numbers:
request_countcounts model calls, not prompts. One prompt can trigger several calls. Don't divide tokens by prompts using this field.- Code review and cloud agent report
product_selected, because the product picks the model. Code review's models report asunknown. So the single biggest unattended consumer in many orgs is exactly the one whose model you can't see. clientis best-effort. It's identified by the integration ID each request sends, so GitHub says to treat it as a grouping, not a verified identity.
This is the closest the metrics API has ever come to cost attribution. Token volume by tool and model is the raw input to a cost model. But it isn't a cost. Per-model rates differ by more than 40x on GitHub's own price list, prompt and output tokens price differently, and cached tokens price differently again. A tool that sends a lot of tokens to a cheap model can cost less than a tool that sends few to an expensive one.
The September charts that were wrong
On October 6, GitHub said that if agent activity looked like it was falling while Copilot usage kept growing, the data was the problem (GitHub Changelog, October 2026). Several IDEs had moved agent sessions onto the Copilot SDK. Those sessions didn't say which IDE they came from, so most of that activity was left out of reports, and some was counted as Copilot CLI.
The fix depends on developers updating. VS Code has it from 1.139.0. Visual Studio, JetBrains, Eclipse and Xcode were due between October and November 2026. Two consequences matter for anyone who reported agent adoption upward:
- The gap can't be backfilled. GitHub says the affected activity doesn't identify its IDE, so it can't be attributed later. Recovery will be gradual, not a single jump.
- Copilot CLI may look inflated. Some SDK-client activity landed there, and earlier CLI numbers can't be corrected.
And the line that matters most for finance: "Billing isn't affected." Agent adoption looked like it was falling while the credits behind it were being billed normally. Anyone who read a dip in agent lines of code as a dip in agent spend got it the wrong way round.
GitHub's suggested check is a good one. Per-user totals_by_ide includes last_known_ide_version and last_known_plugin_version, so you can find who's still on an affected version.
Named skills and MCP servers
On October 22, the totals_by_skill, totals_by_custom_agent, totals_by_mcp, totals_by_slash_cmd and totals_by_plugin arrays started reporting customer-defined items by their configured name instead of other or custom, with up to 25 entries per record instead of five (GitHub Changelog, October 2026).
That's useful for adoption: you can finally see whether the internal MCP server your platform team built gets used. But GitHub labels these metrics client telemetry that "can undercount activity", and says to use them "for trends and adoption rather than for exact accounting". Names only appear from report days processed after the release. They also come from a limited set of clients. IDE integrations not built on the Copilot SDK don't report names at all.
Industry comparisons benchmark adoption, not spend
The same day, the Copilot impact dashboard gained an "Industry comparison" section (GitHub Changelog, October 2026). It's opt-in and off by default. It compares the share of your users who are agent-first or multi-agent, and the pull request merge rate of agentic versus non-agentic users, against peers grouped by company size and one of ten industries.
Notice what isn't on that list. There's no credit consumption, no cost per developer, no spend at all. It's a benchmark of how much you use agents, not of what that use costs. Our earlier piece on benchmarking Copilot spend covers why honest external cost benchmarks are hard to build, and nothing in this release changes that.
The scope notes are tight, too. Comparisons appear only in the dashboard. They aren't in the usage metrics API, the NDJSON exports or downloads. Aggregated benchmarks are deleted after 30 days. Data residency tenants are compared only within their own region. And enterprises used only to assign Copilot Business licenses can't take part.
Our finding: we read GitHub's Copilot metrics reports (the per-user 1-day, organization 28-day and per-repo 1-day reports) but we don't parse
totals_by_clientyet, so FinOpsAid can't currently tell you how much of a user's Copilot traffic came from Claude Code versus VS Code. More important, building it wouldn't give you a cost split on its own. Our credit figures come from GitHub's billing data, where the per-user AI credit number is a single daily total with no client or model breakdown. The metrics API has tokens by client and model, but no credits. Pricing those tokens per model would produce a modeled split, which we'd have to label as modeled and expect to differ from the invoice, because cached-token rates, long-context tiers and promotional prices all sit between a token count and a billed credit. The two ledgers don't join on their own.
What to do with this
- Pull
totals_by_clientbefore you set a tool policy. If you're deciding whether to support OpenCode or Claude Code against Copilot, base it on server-counted request and token volume, not on who complains loudest. - Treat
request_countas model calls. Never divide it by prompts, and don't compare it with interaction counts. - Re-read September's agent charts. If agent activity fell while usage grew, it was probably the SDK undercount. Note the gap in any report that went upward; it can't be backfilled.
- Find outdated IDEs. Use
last_known_ide_versionandlast_known_plugin_versionin per-user reports, and push VS Code to 1.139.0 or later. - Use named skills for adoption, not accounting. They're client telemetry and GitHub says they undercount.
- Opt into industry comparisons only for adoption questions. It won't answer "are we overspending?" and it isn't exportable.
- Reconcile cost from billing, not metrics. When the two disagree, the invoice wins. Our guide to what consumes AI credits explains which features hit the meter.
Where FinOpsAid fits
FinOpsAid reads your org read-only and joins Copilot seat, metrics and billing data with the rest of the GitHub bill. Credits per user, per team and per cost center come from billing. Adoption and activity come from the metrics reports. They're shown side by side rather than blended. Honest limits: the per-user credit figure has no model or client breakdown, we don't parse totals_by_client yet, and GitHub's Copilot metrics lag roughly two to three days. Related: how to read the Copilot page, how to read the benchmarking page, measuring Copilot ROI, and why model routing makes per-model attribution harder in Copilot Fusion billing.
Frequently asked questions
Does the Copilot usage metrics API show Claude Code and OpenCode usage?
Yes, since October 23, 2026. The totals_by_client array reports request counts, session counts and prompt and output tokens per client, including claude_code, opencode, codex and zed, broken down by model. GitHub counts these on the server. It treats the client name as a best-effort grouping based on each request's integration ID.
Can I calculate Copilot cost per tool from totals_by_client?
Only as a model. The array reports tokens, not AI credits, and billing data doesn't break credits down by client or model. To estimate cost per tool you'd price each model's prompt and output tokens at GitHub's per-model rates, and the result won't exactly match the invoice. Cached tokens, long-context tiers and promotions all affect billed credits.
Did the October agent metrics undercount affect my Copilot bill?
No. GitHub says billing wasn't affected. The issue only changed how agent activity from SDK-based IDE versions was attributed in usage metrics, so agent interactions and agent lines of code were undercounted while credits were billed normally. The missing metrics can't be backfilled. They recover as developers update, starting with VS Code 1.139.0.
Sources
- GitHub, "Usage metrics API adds usage from tools like OpenCode", retrieved 2026-11-03, https://github.blog/changelog/2026-10-23-usage-metrics-api-adds-usage-from-tools-like-opencode
- GitHub, "Custom skills, agents, and MCP servers in usage metrics API", retrieved 2026-11-03, https://github.blog/changelog/2026-10-22-custom-skills-agents-and-mcp-servers-in-usage-metrics-api
- GitHub, "Update your IDE to restore agent activity in Copilot usage metrics", retrieved 2026-11-03, https://github.blog/changelog/2026-10-06-update-your-ide-to-restore-agent-activity-in-copilot-usage-metrics
- GitHub, "Industry comparisons in the Copilot impact dashboard", retrieved 2026-11-03, https://github.blog/changelog/2026-10-22-industry-comparisons-in-the-copilot-impact-dashboard
- GitHub Docs, "REST API endpoints for Copilot usage metrics", retrieved 2026-11-03, https://docs.github.com/rest/copilot/copilot-usage-metrics
- GitHub Docs, "Models and pricing for GitHub Copilot", retrieved 2026-11-03, https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing
If you want credits from billing and adoption from metrics side by side, without blending the two, connect your GitHub org — read-only and free during beta — or explore the demo dashboard first.