GitHub Switched Copilot to Token Billing and Developers Are Doing the Math in Public

Copilot circa 2023 was autocomplete. The marginal cost of suggesting the rest of your line was tiny, and $10 a month covered it with room to spare. Copilot in 2026 spins up agents that read your whole repo, run tests in sandboxes, iterate on failures, and open pull requests. A heavy user on a flat plan was costing GitHub orders of magnitude more than a casual one while paying the same $10. Every AI coding vendor has been quietly eating this margin problem for two years. GitHub just blinked first and loudest.

Token billing also has one real virtue: it's honest. You can see exactly what an agent run costs, which has a way of making people suddenly interested in efficient prompts and smaller diffs.

Why developers are still right to be angry

Honest pricing delivered dishonestly still burns trust, and the rollout here was rough.

The complaints filling Reddit and Hacker News mostly aren't "AI should be cheap." They're about predictability. A subscription you can budget; a meter you have to watch. Developers describe kicking off an agent task, getting pulled into a meeting, and coming back to find the agent burned through a week's worth of credits retrying a flaky test. The agent's failure modes are now your bill.

There's also the bait-and-switch texture of it. Teams spent two years wiring Copilot into their workflows at a known price, and the price changed shape after the dependency formed. Business and Enterprise customers get promotional included credits through September 1, which is less a grace period than a countdown.

What to actually do about it

If you're on a Copilot plan, three practical notes from my own week of testing:

1. Check the new usage dashboard before your first bill, not after. The spread between light and heavy agent use is enormous.
2. Scope agent tasks tightly. "Fix this failing test" costs a fraction of "clean up this module." Vague instructions are now literally expensive.
3. Price the alternatives with your real usage numbers, not vibes. Depending on your patterns, flat-rate competitors or cheap API models routed through an open IDE plugin may come out far ahead. Or far behind. The only way to know is the dashboard.

The takeaway

The frustrating thing is that GitHub is probably early, not wrong. Agentic coding has real marginal costs, and the industry's flat-rate era was a customer acquisition subsidy with an expiration date. More vendors will follow; September 1 is the date to watch for how Business customers respond when the promotional credits lapse.

But "we were going to have to charge you eventually" is a reason, not a rollout plan. GitHub turned its most loyal users into meter-watchers overnight, and meter-watchers shop around.