June 11, 2026
The Subsidy Ends
This post is part of my Medium blog.
The day is approaching when the subsidized AI era ends. It hasn't happened yet. But the signals are already visible if you're willing to look.
Pricing gets rational
Anthropic released Claude Fable 5 this week, and the pricing said more than the benchmarks. At $10 per million input tokens and $50 per million output tokens, Fable 5 is exactly twice the price of Opus 4.8, which came out two weeks earlier. That's not a rounding error. That's a statement.
And the messaging around the launch was just as clear: Fable 5 is included on Claude subscriptions until June 22. After that, it goes to usage credits. Pay for what you use. Think about that for a second. The most capable model Anthropic has ever released is included in your subscription for about a week. A week. That's not a subscription benefit. That's a promotion. And most people don't sign up for subscriptions expecting promotions — they expect the thing they're paying for to keep being the thing they're paying for.
This isn't a temporary promotion — it's Anthropic telling you that the most capable model they've ever released doesn't fit inside a flat monthly fee. They watched what happened when OpenAI gave ChatGPT Pro users unlimited access to their most expensive model, and they decided not to repeat that mistake. Expensive models are expensive. Someone has to pay.
And the timing is hard to ignore. Anthropic filed its S-1 earlier this month. OpenAI followed days later. When you're about to go public, you need growth numbers and you need a story about sustainable revenue. Giving everyone a taste of the most powerful model for a week drives sign-ups and engagement metrics. Moving it to pay-as-you-go after that demonstrates pricing discipline to future shareholders. It's promotion for the model and promotion for the offering, at the same time.
This is what rational pricing looks like. The era of models getting cheaper with every release is stalling out, and it's not a choice — it's being forced by constraints that no amount of venture capital can wish away. Energy costs for training and inference are real and growing. Data center power consumption is hitting limits that aren't just financial — they're physical. Inflation has pushed up the cost of everything from chips to cooling. And the supply of capable engineers who can build and operate this infrastructure hasn't kept up with demand.
The economics are catching up. You can subsidize a product for a while, but you can't subsidize the electricity bill forever.
The $200 problem
Consider Cursor Ultra at $200 a month. If you're using it intensively — long sessions, complex refactors, multiple agent loops — you can easily burn $400 or more in token costs in a month. Cursor makes up the difference with venture capital.
Or ChatGPT Pro at $200 a month. OpenAI gives you access to their most capable models with what they call "unlimited" access, except it's not unlimited. There are rate limits, usage caps, and quiet throttling that kicks in when you've consumed more than the subscription economics can justify. And the reports are increasing — more people noticing that quality has degraded, that responses feel different, that the model they're talking to today isn't the same one they were talking to last month. Some of that is model updates. Some of it is throttling. It's hard to tell the difference, which is the point.
Or Claude Max at $100 or $200 a month, depending on the tier. A heavy Claude Max user running extended agent sessions can burn through token allocations that cost far more than the monthly fee. Same pattern: the price doesn't reflect the cost of delivery for the users who actually push the product hard.
Windsurf does it too. Replit's AI tier does it. The various agent platforms that charge a flat monthly fee for what is fundamentally variable, usage-based infrastructure — they're all doing it. Price the product below the cost of delivery, grow the user base, raise another round, repeat. It works beautifully as long as the money keeps flowing.
But the money won't keep flowing forever.
The model shift
The real problem isn't the tools. It's which models people are reaching for.
An entire ecosystem of AI development tools has emerged over the last three years, and most of them default to the most expensive proprietary models — Opus, GPT-4.1, Fable. But look at OpenRouter's leaderboards. Look at what people are actually routing to. The growth isn't in the frontier models from Anthropic and OpenAI. It's in models that don't have nine-figure marketing budgets — DeepSeek, Qwen, Llama, Mistral. Some are open source. Some are from China. Some are both. What they share is that they deliver 80-90% of the capability at a fraction of the cost, and they're not the ones getting the press tour. Just because a model is the one you hear about most doesn't mean it's the one people are using.
This is the shift that matters. Not developers abandoning their tools — you can use Cursor from a CLI, and plenty do. The shift is developers stopping reaching for the most expensive proprietary models when something cheaper gets the job done. The $200/month subscription to a frontier model starts to look different when a model that costs a tenth as much handles 90% of your daily work.
Even Cursor — the company that was burning VC money to subsidize your Opus habit — just shipped Composer 2.5, its own coding model. It matches Opus 4.7 and GPT-5.5 on the benchmarks at roughly a tenth the cost per token. The company that built its business on reselling you expensive proprietary inference is now telling you to use its cheaper model instead. That's not a coincidence. That's the economics catching up.
No, these cheaper models aren't discovering new mathematics independently. But most people don't need a model that discovers new mathematics. They need a model that writes working code, edits existing code, and handles the 90% of daily work that doesn't require frontier reasoning. A model that costs a tenth as much and gets the job done isn't a compromise. It's just the right tool for the work.
The tool companies built on the assumption that users would keep defaulting to the most expensive option. That assumption is already breaking. When your power users start routing to DeepSeek for routine tasks and only hitting Opus for the hard problems, the token economics flip. The subsidized subscription model depends on heavy usage of the most expensive model. If heavy usage shifts to cheaper models, the subsidy math doesn't work the same way — but neither does the revenue.
Some of these companies will raise prices. Some will degrade the experience — slower responses, cheaper default models, usage caps. Some will try to build their own models, which is a different kind of money pit. And some will just run out of runway.
The VC money that made all of this possible is not a permanent feature of the landscape. It has a shelf life. And it's already expiring.
The protectionism begins
There's another signal, and it's not about pricing. It's about fear.
Watch what happens every time a capable model emerges from outside the US. The immediate reaction isn't curiosity. It's a barrage of questions about safety, about data handling, about whether the model can be trusted. Congressional letters. Regulatory inquiries. Op-eds from think tanks that happen to be funded by the companies standing to lose the most.
DeepSeek V4 dropped and the response was instructive. The model is genuinely capable — competitive with frontier offerings at a fraction of the cost, trained on a budget that would barely cover the catering at an OpenAI all-hands. That's the real story. A Chinese lab built something that works, did it cheaply, and released it openly. If DeepSeek can deliver frontier-class inference at a fraction of the cost, the US providers can't hide behind "inference is expensive" forever.
But instead of engaging with the capability, the conversation immediately shifted to whether it was safe, whether it was stealing data, whether it should be restricted. Every time a model comes from outside the US, the same people show up with the same claim: they must have stolen the training data. It's a reflex at this point.
Except distillation is how most model providers operate. Elon Musk admitted to distilling Grok from OpenAI's output. Everyone trains on web content, and a huge and growing share of web content has already been through another model. In some sense, they're all borrowing from each other's output. The "they stole our data" argument isn't wrong because it's unfair — it's wrong because it describes what every provider is already doing. And the content creators who've been saying "you've been using our content all along" have a point that applies to everyone, not just the labs in Beijing.
Some of those safety questions are legitimate. Most of them are not new — they're the same questions that never seemed urgent when the model in question came from San Francisco.
This is protectionism dressed up as prudence. The US AI industry has a growing competitive problem, and the regulatory apparatus is being deployed to slow down the alternatives. When a US model hallucinates, it's a known limitation. When a Chinese model hallucinates, it's a national security concern. The double standard is the tell. Anytime a model shows up from a country that isn't yours, you're going to hear fear, uncertainty, and doubt. That's not a safety posture. It's a market strategy.
The subsidy ends one way. The protectionism is how the incumbents try to extend it another way. Both are temporary. The capabilities are not.
What happens when the subsidy ends
When the subsidy ends, a few things happen simultaneously:
Model prices rise, or quality drops, or both. The providers have to cover their costs. There is no magic.
Tool prices rise, or features get cut, or both. The companies built on subsidized model access have to pass costs through to users or absorb them, and absorbing them requires revenue that most of them don't have yet.
The "AI will be cheap" narrative collapses. For three years, people have been saying that model costs will drop to zero. They won't. Inference is expensive. Training is expensive. The hardware is expensive. The electricity is expensive. The only thing that made it seem cheap was that someone else was paying the bill.
The next subsidy
Here's where this gets interesting. The same people who told you AI would be cheap are the ones who'll tell you AI will fund universal basic income... or will they? The argument goes: AI creates enormous productivity gains, those gains generate enormous wealth, and that wealth gets redistributed through UBI. Clean story. Except the wealth part depends on the cheap part, and the cheap part is already coming apart.
At some point in the future, the conversation about who owns AI's output is going to get serious. Public ownership of AI infrastructure — data centers, models, the compute layer — is already being discussed in policy circles with Bernie Sanders suggesting that the government should own 50% of AI providers. When and if that conversation connects to the UBI conversation and job displacement, you'll have a different kind of subsidy. Not venture capital making your Cursor subscription cheap. Public investment making AI access a utility.
That's a long way off. But the trajectory is visible. The current subsidy ends because VC money tightens and physics catches up. The next subsidy, if it comes, will be political rather than financial. And the interesting question isn't whether AI gets subsidized again. It's what rational pricing looks like when you have to account for both the cost of running the machines and the cost of supporting a society that might be disrupted.
In Essential, the third book in The Condition Set trilogy, the law that requires a human to remain in the decision loop is ninety days from expiring. Rob Coleman runs the agency that oversees every AI system in Canada. The question isn't whether the machines work. The question is whether anyone can tell when they stop working in the public's interest.