Skip to main content
  1. Articles/

GitHub Copilot's Billing Switch Is Live: Here's What Day One Looks Like

·1026 words·5 mins·
Author
Florent Clairambault
CTO & Software engineer

The meter started running at midnight.

As of June 1, 2026, GitHub Copilot no longer sells you a flat monthly subscription. Every plan — Free, Pro, Pro+, Business, Enterprise — now runs on AI Credits: a consumption model where you pay per token, per model, per call. The flat-rate era that made Copilot a safe “fixed cost” line in every engineering budget is over.

Developers had six weeks of warning. It wasn’t enough.

What Actually Changed
#

The mechanics are straightforward, but the implications are not.

One AI Credit equals $0.01. Credits are consumed based on token usage: input tokens, output tokens, and cached tokens, billed at the published API rate for whichever model you’re using. The models are not priced equally. Claude Opus 4.7, for instance, carries a 27x credit multiplier relative to the cheapest available model. When the Billing Preview tool rolled out in April, one developer running it against their actual April usage saw a $39 estimate balloon to $902. That number made the rounds. It was not an outlier.

Here’s what didn’t change: code completions and Next Edit suggestions remain free. The core autocomplete experience — the feature that made Copilot famous in 2021 — doesn’t consume AI Credits. If that’s all you use, June 1 is a non-event.

But if you use chat, run agents, rely on code review, or do any of the agentic workflows that GitHub spent the last year promoting, the meter is running.

The Developer Reaction
#

The official GitHub community announcement thread collected 893 downvotes and more than 400 comments before the billing switch went live. That number is not typical. For context, contentious GitHub changelog posts rarely break triple digits.

The concerns are not philosophical. They’re arithmetic.

A solo developer who runs agentic coding sessions — the kind where Copilot orchestrates multi-step file changes across a repository — estimated routine sessions consuming $30–$40 each. At the Pro plan’s included credit balance, that’s one working session before the meter starts adding to the bill. A team that uses Copilot’s code review feature at scale discovered they would be billed twice: once in AI Credits for the review itself, once in GitHub Actions minutes for the CI runner that executes it. The double-billing problem drew particular anger.

Specific numbers circulating in developer forums: a developer who was paying $29/month projects their team’s new bill at $750. Another projects a jump from $50 to $3,000. These aren’t edge cases — they’re engineers who use Copilot the way GitHub has been telling them to use it since Copilot X launched.

The Control You Have
#

GitHub built hard cap controls into the new system, and this is worth emphasizing: if you set your spending cap to $0, Copilot will pause when your included credits run out rather than charge you. It won’t silently accumulate a bill. It stops.

The worst case for an inattentive user is that Copilot becomes unavailable for the rest of the billing period — annoying, not catastrophic. For solo developers, setting the cap at $0 is almost certainly the right default this month while you understand your actual usage pattern.

Organizations need to move faster. Enterprise and Business plans have per-seat credit budgets and can set per-user caps. If your engineering manager hasn’t reviewed those settings today, they should.

The Billing Preview tool remains available after the switch. You can project next month’s cost against current usage before you commit to any cap level.

Who Gets Hurt Most
#

Not all Copilot users are equally exposed.

Heavy chat users hit the model premium first. GPT-4.1 and Claude Opus 4.7 are the models most developers reach for in chat. Both carry significant multipliers compared to the cheaper models.

Code review users face the double-billing problem. If your workflow involves auto-triggering Copilot code review on every PR, you’re now paying AI Credits plus Actions minutes simultaneously.

Agentic workflow users face the steepest exposure. Multi-step Copilot agents consume tokens at a rate that’s an order of magnitude higher than chat. The same session that used to fall inside a flat-rate cap now itemizes every tool call.

Light autocomplete users are essentially unaffected. Completions are free. If Copilot is a tab-completion layer for you, nothing changed.

Where Developers Are Going
#

The community discussion threads are showing a consistent pattern: developers are actively shopping alternatives.

The most frequently cited destinations are Claude Code (on the Max plan, which remains flat-rate at $100 or $200/month regardless of usage intensity), DeepSeek V4 (open-weight, $1.74/M input tokens, self-hostable), and local models via Ollama. The irony of developers switching to Anthropic’s product is not lost on anyone — Claude Opus 4.7 is one of the models that makes Copilot expensive, but accessed directly through Claude Code, it’s covered under a fixed subscription rather than billed per token.

GitHub suspended new individual and student signups in the lead-up to the switch, citing system preparation. That’s an unusual move that has added to the sense that Microsoft expected resistance.

The Structural Shift
#

What happened today is not a pricing change. It’s a category change.

GitHub Copilot launched as a productivity tool — a fixed cost you bought once and forgot about, like a JetBrains subscription. Somewhere between Copilot Chat, Copilot Agents, and the enterprise feature buildout, it became infrastructure: a service whose cost scales with how much work you put through it.

That’s a legitimate model for infrastructure. AWS doesn’t charge you a flat rate for EC2. But AWS also never marketed EC2 as a flat-rate product and then switched the billing model after developers had built workflows around it.

The flat-rate era trained developers to use Copilot aggressively. The consumption era arrived without those habits changing first.

Set your cap today. Review your usage data. And if the numbers don’t work at your current workflow intensity, the alternatives have never been more capable.


Sources:

Related