It happened on the date. On June 1, 2026, GitHub moved every Copilot plan onto usage-based AI Credits. Code review now consumes Actions minutes. There are user-level budgets to cap spend, and a new Copilot Max tier for anyone who burns through the included allowance.
A few months back, in the Copilot tokenizes trilogy, I argued the leaked June 1 date was the tell: a flat-rate assistant quietly becoming a metered one. The invoice arrived on schedule. This is the receipt, and the part that actually matters now.
What Flipped
- Every plan is metered. AI Credits are the unit. Your bill is no longer a subscription line, it’s a consumption line.
- Code review costs Actions minutes. A feature that read as “included” now draws down a budget you were already watching for CI.
- Budgets and Max. User-level budgets exist because spend is now variable and someone has to govern it. Copilot Max exists because the meter, left alone, runs past the old flat fee.
Predictable Became Variable
Copilot went from “costs $X a month” to “costs whatever you used.” Flat-rate has one property metering destroys: you can budget it. A team of forty on a fixed seat is a known number. On AI Credits it’s a distribution with a long tail, and the tail is your heaviest users.
Don’t call that “punishing your best people.” Tokens are an input, and heavy usage is not high performance. A sharp engineer who’s economical with prompts barely registers; someone burning credits on slop tops the chart. Metering taxes consumption, blind to whether it shipped anything. If a heavy user’s tokens convert to value, paying in proportion is fair.
The real damage is legibility: a cost you can see always beats a value you can’t. Put token spend on a dashboard and you’ve handed every budget owner a hit list sorted by the wrong metric. Your highest-leverage engineers top a “who’s expensive” report with no “who’s valuable” report beside it. Finance optimizes what’s legible, and the only legible number is the bill. So scrutiny and budget caps land on the people generating the most leverage, and eventually someone’s review cites their AI spend as if it were a finding. The meter doesn’t just measure your best people. It paints a target on them.
— A developer reacting to the changeWe adopted it because it felt free at the margin. Now every suggestion has a price, and you can feel yourself hesitating before you ask.
That hesitation is the product change. A tool you reach for without thinking and a tool you meter are not the same tool.
The headline is AI Credits; the quieter line is code review eating Actions minutes. Bundled features turning into separately-metered ones is the move to watch across every vendor, because it never reads as a price increase. Audit what used to be “included” before the first variable invoice, not after.
The Review Tax
There’s a second-order effect worse than the budgeting one. Code review now burns Actions minutes, and review is exactly where tokens pile up: re-reading diffs, reasoning about edge cases, checking the thing the model confidently got wrong. Metering quietly puts a price tag on the one step that catches mistakes.
So the incentive inverts. Generation stays cheap to reach for; thorough review now shows up on the bill. Teams under pressure do what metered incentives always produce: generate freely, review sparingly. And it lands at the worst moment, with AI shipping more code than any reviewer can keep up with. Meter the review on top of that and the gap fills with unreviewed AI code that merely looks done.
The quality hit is invisible until it isn’t. Nobody files a ticket called “we reviewed less this quarter.” It surfaces later as defects, regressions, and the 44 cents of every dollar spent fixing what shipped wrong. Metering doesn’t just risk taxing your best people. It taxes the verification that keeps everyone else’s output safe.
Why It’s Also Reasonable
I’m skeptical by default, so credit where it’s due:
- Inference costs real money. Flat-rate never survives heavy agentic use, and subsidizing it forever isn’t a business.
- Budgets are a real feature. User-level caps give teams a brake flat-rate never offered.
- It’s the whole industry. The same current runs through Anthropic’s enterprise tax and the broader margin story. Metered-by-default is where agentic tooling is settling.
The Takeaway
“Usage-based” is the polite phrase for “the bill now scales with how useful the tool is.” Defensible economics, and a real change in how you plan: you budget differently, you govern differently, and your heaviest users feel watched.
The trilogy was about the date. The date came. Now make sure you’re measuring consumption before it’s measuring you.
Related Posts
