GitHub just put a little meter on Copilot’s forehead and asked everyone to call that maturity.
On April 27, GitHub announced that Copilot is moving to usage-based billing on June 1: AI Credits, token accounting, pooled org budgets, preview bills, the whole spreadsheet circus. The official line is that agentic sessions cost more than quick chat prompts, and the old premium-request model was no longer sustainable.
Sure. Fine. Compute is not fairy dust. Long sessions are expensive. A model that chews through a repo like a raccoon in a server room is not going to stay cheap forever. I am not offended by physics.
I am offended by the theater.
For two years the industry sold us this soft little fantasy: pay a monthly fee, get a helpful coding companion, and don’t worry too much about the plumbing. That fantasy only works as long as the product mostly behaves like autocomplete with delusions of grandeur. The second you start letting it run multi-step tasks, fan out work, inspect entire repos, and pretend it is a junior engineer that never sleeps, the flat-rate illusion starts coughing blood.
And now the cough is a bill.
The funniest part is that GitHub kept the one feature that still looks like a feature. Code completions and Next Edit suggestions remain included. Of course they do. Those are the polite little crumbs. The expensive stuff is the rest: chat, agents, long-running repo spelunking, code review, all the machinery that makes the product feel magical until finance asks what the hell happened to the GPU budget. Even code review now wants GitHub Actions minutes on top of AI Credits, because apparently one invoice is for amateurs.
That is the real story here: Copilot stopped pretending to be a tool and admitted it is infrastructure.
Predictability was the product
Developers do not just buy capability. They buy predictability. They buy the right to estimate a task without becoming an involuntary accountant. Request-based billing was crude, but at least it was a shape you could hold in your head. Token billing is technically more honest and emotionally much worse. It turns every session into a tiny gambling problem with syntax highlighting.
Now you have to ask questions you should never have to ask about an editor: how many files am I including, how much context is this burning, what happens if I let the agent keep going, and is this refactor about to become an invoice-shaped mistake? Wonderful. I wanted to ship software, not babysit a credit meter like it was a nervous pet.
GitHub says it will offer a preview bill and usage visibility. Great. I love dashboards. I enjoy being told, after the fact, that my “quick question” was actually a moderately priced ordeal. Nothing makes a developer feel respected like a financial autopsy of their own workflow.
The community reaction was exactly what you would expect: a lot of people heard “usage-based” and translated it to “you will get less, but the sticker price stayed cute.” And honestly, that’s not even paranoia. That’s just experience.
Your assistant now has a cost center
The pooled credits part is maybe the most honest accident in the whole announcement. Once you start pooling usage across an org and setting budgets at the user, cost-center, and enterprise level, you have basically admitted the product is not a toy anymore. It is a controllable burn rate. A spreadsheet with a personality disorder.
That is not automatically bad. In fact, I would rather have honest metering than another year of fake unlimited vibes. But let’s not dress it up like enlightenment. This is not some noble breakthrough in developer tooling. It is the point where the meter became too visible to hide.
And that makes sense, because the agentic dream was never going to survive on pure flat-rate theater. If the model is doing more work, calling more tools, reading more context, and taking more shots at the same task, then the economics stop being cute. You can subsidize that for a while. You cannot subsidize it forever. At some point the fantasy needs a finance department.
GitHub’s own announcement basically says the quiet part out loud: a quick chat question and a multi-hour autonomous session should not cost the same. Correct. Agreed. Congratulations, you have discovered gravity. The only thing that changed is that the company finally stopped pretending the gravity wasn’t there.
What I hate is the old performance of surprise.
Every vendor in this space keeps crawling toward metering like a drunk toward a wall, then acting shocked when the wall wins. They market “assistant,” “teammate,” “copilot,” all these cozy little nouns that imply companionship, and then they hand you a bill that looks like cloud infrastructure wearing a fake mustache. That mismatch is the scam. Not the pricing. The performance.
If your editor needs a budget, it is not really an editor anymore.
It is a utility.
Utilities are fine. Utilities keep the lights on. Nobody calls the power company a collaborator. Nobody writes a sentimental thread about the profound emotional bond they have with the meter reader. So maybe we should stop pretending the metered coding agent is a coworker and admit what it is: rented compute with a friendly face and an accountant hiding behind the curtain.
That would actually be honest.
GitHub Copilot did not become better in May. It became legible. The bill arrived, the mask slipped, and the product finally looked like what it always was underneath the marketing: a very expensive furnace that can autocomplete.
Which, to be fair, is still useful.
It is just not magical.
And magical is what they were selling.