github-copilot
GitHub Copilot
Concept of the skill
What it is: A routing-and-cost discipline for GitHub Copilot — knowing when its metered budget is worth spending on a task, how its billing works (premium requests with per-model multipliers, replaced June 1 2026 by usage-based AI credits at 1 credit = $0.01), what surfaces are always free, and when a cheaper or free lane should take the work instead.
Mental model: Copilot has two cost surfaces glued together: an always-free completions/next-edit surface (unmetered) and a metered chat/agent surface (premium requests → AI credits). Routing well means keeping low-complexity and high-volume work off the metered surface and reserving the finite, non-rolling metered allowance for work that genuinely benefits from an IDE-native frontier model.
Why it exists: The metered allowance is finite and expires each cycle (no rollover), and agent-like flows burn it fastest — so spending it on boilerplate or work a free lane handles is pure waste. A deliberate routing rule prevents that waste while still using Copilot where its IDE-native flow earns the spend.
What it is NOT: It is not a model and not a runtime — it is a metered access lane to frontier models inside the IDE. It is not "unlimited because I pay a subscription": the subscription buys completions plus a capped metered allowance, not unlimited agent flows.
Adjacent concepts: the metered allowance (the finite budget); the always-free completion surface; the per-model token rate that converts usage to credits; the cheaper/free lane that should absorb low-complexity work.
One-line analogy: Copilot is a prepaid toll lane bundled with a free service road — the service road (completions) is unlimited, the toll lane (agent/chat credits) is fast but metered and the balance does not roll over, so you only take the toll lane when the trip is worth it.
Common misconception: That "I pay for Copilot, so using it is free/unlimited," or that the old per-model multipliers still govern cost. Neither holds post-June-1-2026: cost is token-based (input + output + cached × per-model rate → credits), the allowance is finite and non-rolling, and agent flows consume it fastest.