Why Your AI Subscription Feels Different Now
The truth about the AI buffet model that was never meant to last
The cheap AI era is coming to an end.
I believe most people are already aware, but let me try to explain it in the simplest way possible.
Picture your favourite all-you-can-eat buffet. The first time you go, it is incredible. Endless plates. One low price. You go back the next week. You bring your friends. Your friends bring their friends. The place is packed every night, and the food just keeps coming.
Then one day you walk in and the manager is standing at the door with a slightly uncomfortable expression. The crab legs are gone. The prawns have a serving limit. There is a new premium tier for the good stuff. The sign on the window still says all-you-can-eat, but something has clearly changed.
This is exactly what is happening with AI right now.
For the past two years, using AI felt like a magical, unlimited buffet. You could chat, generate, analyse, summarise, code, and create as much as you wanted for a flat monthly fee. The food kept coming. And behind the scenes, the restaurant was losing money on every plate.
That era is ending. And the reason it is ending tells you a lot about where AI is heading next.
How the AI Buffet Got Started
To understand why the buffet is changing, it helps to understand how it was funded in the first place.
When OpenAI launched ChatGPT Plus in 2023, the pricing was deliberately attractive. The goal was not to make money immediately. It was to get as many people as possible using the product, so they would form habits, build workflows, and become the kind of customer who would pay more later.
Think of it like Uber offering heavily discounted rides in a new city, or Costco handing out free samples at every corner of the store. The price is a customer acquisition strategy, not a reflection of actual costs.
It worked extraordinarily well. ChatGPT reached 100 million daily active users faster than any consumer app in history. Anthropic’s revenue grew 44 times in just 15 months. Millions of people signed up for flat-rate subscriptions and started using AI for everything.
The economics were sustainable because of how people naturally use AI in conversation. You type a message. You read the response. You think. Maybe you type another message five minutes later. The AI is only actively working during those brief exchanges. The rest of the time, the server is idle. Flat-rate subscriptions work when usage is spread out and human-paced, because you can average the costs across millions of people who use it at different times.
This is the buffet logic: nobody eats non-stop. Most customers take a plate, sit down, eat slowly, and leave. The restaurant can price for the average customer because the outliers do not break the model.
Then came the agents.
AI Agents: The Customers Who Never Left
An AI agent is not like a chatbot. You do not have a conversation with it and then go about your day. You give it a goal, and it works toward that goal continuously, taking actions, checking results, correcting mistakes, running searches, writing code, sending requests, and retrying when something fails. Over and over, without stopping to wait for you.
In buffet terms, the other customers left after dinner. The agents stayed. And they kept eating all night.
This changes the economics completely. One agent working on a complex software project for a day might make thousands of API calls. Each call consumes compute, which costs money.
The flat monthly subscription that made sense for a human having a few conversations per day makes very little sense for an agent running thousands of operations per hour.
The numbers bear this out in uncomfortable ways. One developer’s coding project with Claude ran up costs exceeding $150,000 in a single month through API usage. A power user documented by The New York Times burned through 1.8 million tokens in a single month, racking up a bill of $3,600. These are not bugs or misconfigurations. They are a preview of what normal agentic usage looks like at the infrastructure level.
As Eric Broda wrote in his Substack piece “The End of Token Subsidies”:
“For two years, companies treated subsidised AI pricing as if it were durable economics. It was not. It was promotional pricing for a new kind of compute habit.”
That promotional period is over. The kitchen is starting to ask questions about where all the food went.
Who Was Paying for All of It?
The honest answer is: mostly investors.
OpenAI has raised over $17.9 billion and is not yet consistently profitable despite enormous revenue. Anthropic has raised billions on the premise that it will eventually capture enough of the market to justify the losses.
The data centre buildout required to run these models at scale costs tens of billions of dollars in capital expenditure. Someone is funding the gap between what AI costs to run and what users actually pay.
For now, that gap is funded by venture capital, tech giants with deep pockets, and increasingly by sovereign wealth funds and government-adjacent investors who see AI infrastructure as strategic. Governments have indirectly contributed through tax incentives for data centre construction and energy infrastructure.
This is not a criticism. It is how new technology industries work. The telephone network was not profitable for decades. The internet required massive subsidies before it became commercially self-sustaining.
The subsidies create adoption, adoption creates data and capability, and capability creates a product worth paying a real price for.
But subsidies have to end eventually. Investors want returns. The losses cannot grow indefinitely. And the arrival of agents has dramatically accelerated the timeline on which that reckoning arrives.
What the Buffet Looks Like Now
The changes are already underway, even if they are not always announced in big press releases.
Anthropic quietly tested removing Claude Code from its standard $20 per month Pro plan in early 2026. Claude’s Max plan now costs $200 per month and is positioned explicitly for heavy users.
OpenAI has introduced usage limits on its most capable models and launched higher tiers for users who need more. The era of genuinely unlimited access at a fixed low price is narrowing.
The direction of travel is clear. AI is moving toward usage-based billing models where heavy consumers pay proportionally more. There will be different tiers for access, with various features costing different amounts. Users will have better tools to monitor and manage their AI spending before receiving their bills.
Is this bad news? Not necessarily.
When companies are forced to operate more efficiently, they tend to innovate. Smaller, smarter AI models are already appearing that can do specific jobs at a fraction of the cost of the big general-purpose ones. The pressure to be profitable is pushing the industry toward more sustainable engineering choices.
But there’s no getting around the other part of the picture either. As Ed Zitron's analysis pointed out:more compute does not make OpenAI or Anthropic’s services cheaper to offer. The path to profitability goes through higher prices.
The buffet is switching to “pay by the plate.” For light, occasional users, that might not change much. For heavy users and businesses running autonomous agents around the clock, the bill is about to look very different.
What This Means for You
If you are an individual user, the practical impact depends on how you use AI.
Casual users who chat with AI occasionally and use it for occasional writing, research, or creative projects will probably not notice much difference. Standard tiers will continue to exist at reasonable prices. The changes are aimed at the heavy end of usage.
If you are a developer or small business owner who is starting to build with agents, this is the moment to start paying attention to your AI costs the same way you pay attention to your cloud infrastructure costs. Set budgets. Monitor usage. Build with efficient models where the task does not require the most capable option available. The habit of treating AI spending as a variable cost to be managed rather than a flat subscription to be ignored is one worth developing now, before it becomes urgent.
For businesses already running significant AI workloads, the advice is to audit what your agents are actually doing, identify where compute is being consumed unnecessarily, and build approval workflows for tasks that might trigger large unexpected bills.
For everyone: AI is not going away. It is growing up. It is moving from a magical technology subsidised by venture capital into something more like a real utility, with real bills, real pricing tiers, and real consequences for wasteful usage. This is what maturity looks like in a technology industry.
The Buffet Will Still Be Open
The unlimited era was genuinely fun while it lasted. Cheap, powerful, always-on AI changed how millions of people work, create, and solve problems. That shift is permanent regardless of what happens to pricing.
What is changing is the economic model underneath it. The kitchen is becoming more honest about what things actually cost. The all-you-can-eat sign is coming down in some places and being replaced with something more like a good restaurant: excellent food, clear prices, and the expectation that serious eaters will pay a serious bill.
That is not a tragedy. That is sustainability.
The tools will keep improving. The prices will keep coming down per unit of capability. And the companies that figure out how to use AI thoughtfully, measuring what they use and optimising how they use it, will find they are spending less and getting more than the ones who treated the buffet as something that would last forever.
Nothing does. But the kitchen is still open.
Check out Eric Broda’s “The End of Token Subsidies” and Ed Zitron’s “AI Is Too Expensive” on Substack. I recommend these reads for anyone involved in AI infrastructure.