Cheap AI is over. Copilot just stopped pretending
Copilot's transition to usage-based billing looks like boring news from the category of “they changed the tariff.” Formally, everything is really neat: instead of abstract requests there are now credits, instead of sensations there are tokens, input, output, cache. Even the base prices are not touched, so as not to frighten once again.
But this is not about billing. This is about the moment when the industry stops pretending that intelligence can be sold by subscription, like music or movies.
For the last couple of years, the market has lived in a strange structure, it seemed that for a fixed $10–40 per month you get access to something that, by its nature, costs much more. It worked as long as the AI was easy: chat, auto-complete, quick replies. So far it looked like a tool, not a full-fledged workforce.
Copilot price update from June 1
The problem started the moment the agent appeared.
An agent is no longer an “answer to a question.” It's a process. Long, repetitive, with memory, with context, with trial and error. He reads the code, keeps the project structure in his head, rewrites something, checks, rolls back, tries again. That is, it begins to behave not like a function, but like a developer. And, like any developer, he works not just once, but constantly.
From an economic point of view, this breaks everything. Because this is not one request, but hundreds and thousands of small operations under the hood. And each of them is a calculation.
The most unpleasant thing about this story is that the breakup has already happened. A real agent session today easily burns tens of dollars per working day. Not because someone is “using it incorrectly,” but because the model itself is like this: a large context, many steps, constant iterations. The more serious the task, the more expensive it is.
And at this moment a conflict between two realities arises. One Copilot costs $10. In another, it is a tool that is closer in load to a cloud server than to a plugin in the editor. GitHub simply stopped keeping this difference at its own expense.
There is hope that all this is temporary, that models will become cheaper, hardware will speed up, and eventually we will come to almost free intelligence. This sounds logical until the first real use. Because every time something becomes cheaper, the market immediately begins to use it more deeply. The context becomes longer, the tasks are more complex, the expectations are higher. You don't start spending less - you start doing more. And the score remains.
The main myth of recent years has been almost childish: that somewhere in a data center a “free country of geniuses” has appeared, which can be rented by subscription. In reality, there are simply very expensive machines that can imitate intelligence. And no one is going to give these cars away for free indefinitely.
Copilot has not become worse in this sense
This whole story is unpleasant because it changes the economics of development not towards cheaper, but towards more complexity. AI doesn't remove value, it redistributes it. Some of the money goes from salaries and teams into computing, APIs and infrastructure. A good AI specialist can really do things faster and cheaper. Sometimes much cheaper. But this does not turn complex products into free ones. It just changes where exactly the money goes.
Previously, getting into development was primitively simple. You might have an old laptop from your parents and plenty of time to figure it out. You sat, read, broke down, rewrote, and gradually gathered understanding. It was a long journey, but it required almost no money. The main resource is time and stubbornness.
Now this scenario is starting to break down. The entry threshold has not dropped. He has grown up. And in two dimensions at once: knowledge and money.
On the one hand, you no longer need to deeply understand HTML, CSS, architecture, or the principles of building systems. You can simply tell the agent: “do it.” And he will actually do something. At the interface level this looks like a radical simplification.
On the other hand, without understanding you very quickly begin to burn money into thin air.
Because an agent is not a performer in the classical sense. It is your thinking amplifier. If there is no thinking, it begins to intensify the delirium. You formulate the problem crookedly, he does it crookedly, you don’t understand where the mistake is, he makes another go, adds crutches, increases the context, spends more tokens. A few hours later you are already minus $50–100, and the result still does not work.
And at some point a rather humiliating dialogue occurs:
“I did what you asked, but it doesn’t work”
- and it’s true.
Not because the model is bad. But because you don’t understand what exactly you are asking for and how it should be arranged.
As a result, a new situation arises. Formally, entry into development has become easier. In fact, it’s more expensive and more complicated.
If you used to pay with time for learning, now you pay with both time and money for every attempt. The error is no longer free. Misunderstanding is no longer free. Even research starts to cost.
The gap between markets becomes even more noticeable
For those who have money, this is an accelerator. They get through the error cycle faster because they can afford to burn hundreds of dollars in the process. For others it is a limiter. Any long attempt becomes a risk rather than a natural part of learning.
The myth of “AI for everyone” begins to crack at this point. Everyone has access. The ability to use it effectively is no longer there.
As a result, a familiar thing happens. The technology that was supposed to level the field begins to stratify it. Strong specialists accelerate because they understand what they are doing and how to formulate tasks. The rest get the illusion that they are “developers too” until they are faced with the real cost and real complexity.
Copilot is simply the first here to stop pretending that it can be hidden behind a subscription. Everyone else will come for him. Because AI has no other stable model.
And if you remove all the corporate language, the picture turns out to be quite straightforward. There will be no cheap development. There will be no free development. There will be a faster, more powerful, but also more demanding environment where mistakes cost money and misunderstandings very quickly turn into bills.
Because behind every “smart agent” there is not a genius, but a system that honestly counts your every unsuccessful attempt.
Loading comments...