The Buffet Is Closing: Your AI Subscription Is Becoming a Utility Bill

On June 1, GitHub Copilot (Microsoft's coding assistant, the most widely used AI programming tool in the world) changed how it charges. It used to be a flat monthly fee, the same price whether you used it a little or a lot. Now it bills by usage (you pay for what you consume, like a water or electricity meter). Each plan comes with a monthly allotment: the Pro plan is $10 with $10 of "AI Credits," Pro+ is $39 with $39, the enterprise tier is also $39 with $39. Burn through your credits and you pay more to keep going. Anything left over at month's end does not roll over. It expires.

The reaction in the official discussion thread was almost uniformly angry. The line that stuck with me: one user let the AI run four tasks on its own and watched it eat 53% of a 7,000-credit monthly allotment in a single day, on pace to hit zero in two. Another did the math and figured the same workload could cost "50 times" what it used to.

Most people read this as one word: a price hike. The hike is the surface. The deeper shift, the one worth everyone's attention, is that the billing logic itself changed, from an all-you-can-eat buffet to a running meter. What really separates these two models is where the risk lands. Under the buffet, you pay a fixed price and the platform eats the overage. Under the meter, the harder you push, the bigger your bill, and the cost risk moves entirely onto you.

Why was the change unavoidable? Because how people use these tools changed this year. You used to ask one question and get one answer, which costs very little. The headline feature now is the agent (a system that runs many steps on its own, reading files, editing code, checking its own work). You hand it a task and behind the scenes it can burn dozens of times the compute going back and forth. Here is the irony: the platforms spent two years teaching you to let the agent run free, and that trained habit is exactly the most expensive thing you can now do.

This is not a Copilot problem alone. Last June, the hot AI editor Cursor did the very same thing: it quietly shrank its fixed allowance, moved to usage-based pricing, set off a wave of cancellations, drew a public apology from its CEO, and pushed a chunk of loyal users toward a cheaper rival. Two leading products making the same cut within a year tells you this is structural. That early "$20 and use it all you want" phase was a land-grab subsidy paid for with investor money. Subsidies end. Real costs find their way back onto the bill.

So for anyone who relies on AI tools to get work done and has to account for the cost, here is a concrete takeaway: stop treating AI tools as a buy-it-and-forget subscription, and manage them like a utility. They are a variable cost now, in the same family as cloud servers and your power bill, with the same risk of a nasty surprise. When you hand tools to a team, set caps, watch usage, and review who is burning what and where. One unattended agent spinning in the background overnight is a bill you did not plan for, waiting at month's end.

There is a new opportunity hiding in here too. Once every call maps directly to money, "getting the same job done with less compute" becomes, for the first time, a skill that saves real cash. Writing clear instructions, slicing tasks small and precise, knowing when to let the agent loose and when three lines of your own will do, all of this used to look like fussiness. Now it goes straight onto the cost sheet. I run coding agents all day, and the clearest change this month is that I pause to ask "is this task worth a full autonomous run" before I hit go, instead of reaching for it by reflex.

The magic always had a price. That cost used to sit on someone else's tab, so you never felt it, and assumed it was free. Now it sits on yours. That may not be a bad thing. Starting this month, at least, you get to see exactly how much you pay each day for which kind of convenience, and then decide which conveniences are worth buying and which you can just do yourself.