Share with your CTO
GitHub is closing a visibility gap that has quietly grown as developers route Copilot requests through tools outside the IDE. The Copilot usage metrics API now surfaces client-level telemetry for tools like OpenCode, Claude Code, Codex, and Zed, not just VS Code and JetBrains. A new totals_by_client array breaks down request counts, session counts, and token consumption by tool and by model, available at both the enterprise and organization level in 1-day and 28-day reports.
What this means for your business
Before this update, any developer using Copilot as a model backend through a non-IDE terminal agent produced usage that simply disappeared from enterprise reports. Token costs accrued, governance posture eroded, and the people responsible for both had no data to act on. That’s not a minor reporting gap. It’s the kind of blind spot that turns a controlled AI budget into a surprise line item at quarter-end.
The more significant signal here is what the client list reveals about developer behavior. OpenCode, Zed, and Claude Code appearing as named, trackable clients means GitHub is acknowledging that Copilot has become an API layer, not just an IDE plugin. Developers are composing their own toolchains and routing enterprise model access through whatever interface fits their workflow. The new model_selection_mode field, which distinguishes between user-selected models, auto-selected, and product-selected, makes it possible to see whether developers are actively overriding default models, which is a governance signal most CTOs currently have no way to capture.
The question worth holding: if this data shows that 30 percent of your Copilot token volume flows through unmanaged third-party clients, what’s your policy response? The API gives you the measurement. The harder work is deciding which clients are sanctioned, which trigger a procurement conversation, and which get blocked at the identity layer. Measurement without a governance framework attached is just a more detailed invoice.
Concept deep-dive: Model selection mode
Model selection mode describes how a given AI model was chosen for a request: by the user explicitly picking it, by an automatic routing system, or by the product itself. It exists because the same Copilot subscription can serve requests to GPT-4o, Claude 3.5 Sonnet, or other models depending on what the calling tool requests. Think of it as knowing not just which highway cars took, but whether the driver chose the route or the GPS did. For CTOs, it surfaces whether developers are actively shopping for more capable or cheaper models outside sanctioned defaults.
Based on reporting from Usage metrics API adds usage from tools like OpenCode, originally published 2026-10-23 19:42:00.
