The frontier model subscription was always a lie.

Here’s the timeline.

On June 9, Anthropic launched Claude Fable 5 – a public-facing version of Mythos, the model Washington was already nervous about. Pro, Max, Team, and Enterprise subscribers got access at no extra cost. The window was two weeks. After June 23, Fable would move to usage-based credits at $10 per million input tokens and $50 per million output tokens. Anthropic said it would restore subscription access “when sufficient capacity allows” (Claude Fable 5 and Mythos 5 Launch, Anthropic, June 9, 2026).

WHY FABLE DISAPPEARED

Three days later, the U.S. government made the question moot. Anthropic said it received a Commerce Department directive restricting Fable 5 and Mythos 5 under export controls, citing national security concerns tied to the models’ cyber capabilities. The letter arrived at 5:21 p.m. ET on a Friday. Anthropic disabled both models entirely – not because the order required a full shutdown, but because the company said it couldn’t surgically restrict foreign-national access on the timeline given (AP News, June 13, 2026).

So Fable lasted three days in the wild. The export-control drama is real – frontier models are now being treated as strategic assets, not just software features. But it risks obscuring the more important point: Anthropic had already told us, before Washington stepped in, that flat-rate access to its best public model was temporary by design.

The subscription couldn’t hold

Anthropic called this a capacity issue. Fine. Capacity is economics with a nicer label.

If your best model only fits inside Pro and Max for a two-week promotional window before being carved into usage credits, the flat-rate subscription has already failed as a packaging model.

THE REAL PRODUCT

If your best model only fits inside a flat subscription for a two-week promotional window, the real offering was never the subscription. It was metered compute with an on-ramp.

Functionally, the real offering was a metered product with a promotional subscription window.

And Anthropic is not alone. GitHub moved all Copilot plans to usage-based billing on June 1, with Claude Opus 4.8 restricted to Pro+, Business, and Enterprise tiers and carrying a 15x premium request multiplier (GitHub Blog, June 1, 2026). Two different companies. Same move: frontier models gated behind premium tiers, weighted with multipliers, and funneled toward metered billing. The premium tiers are not flat-rate entitlements anymore. They are access-control layers for scarce compute.

But this is the symptom, not the disease. The question is why the compute economics are this unstable in the first place.

The budget model is breaking on both sides

Uber burned through its entire annual AI coding budget in four months and imposed a $1,500 monthly cap per employee and per agentic tool (TechCrunch, June 2, 2026). Goldman Sachs interviews with 40 software companies found many overrunning initial AI budgets “by orders of magnitude,” with some spending as much as 10 percent of total engineering labor costs on AI tools (The Atlantic, May 2026). Priceline reportedly saw its Cursor renewal come back 4–5x more expensive (TechCrunch, June 5, 2026).

Stanford’s Digital Economy Lab helps explain why. Agentic coding tasks consume roughly 1,000 times more tokens than chat-based coding assistance. The same task can vary by up to 30x between runs. And higher token consumption does not translate into higher accuracy (Stanford Digital Economy Lab, 2026).

Expensive software is manageable. Volatile software isn’t. CFOs can plan for a known bill. They cannot plan for a cost function that spikes 30x on the same task without producing better results. Dario Amodei put the compute problem starkly in February: if an AI lab overbuys compute and the revenue forecast misses, “there’s no hedge on Earth” that saves it from going bankrupt (Fortune, February 14, 2026).

Agentic coding is not a SaaS seat problem. It is a cloud-cost-management problem wearing a chat UI.

But normal cloud costs don’t behave like this either. AWS doesn’t get more expensive next month because a software layer upstream got more verbose or less efficient. Agentic AI can.

That is the demand-side story. But there may be a supply-side story too.

The builders are using their own tools

Anthropic says that as of May 2026, more than 80 percent of the code merged into its codebase was authored by Claude. Engineers are merging eight times as much code per day as they were in 2024. The company describes this as Claude authoring the code while engineers direct and review (Anthropic, ‘When AI builds itself’, 2026).

Eighty percent of merged code. Eight times the daily volume. And Anthropic is not the only lab doing this – it is simply the one that published the number.

Now connect the dots – carefully.

We cannot prove from public evidence that this volume of AI-authored code is making Anthropic’s infrastructure less efficient. But the ingredients are worth naming. Eighty percent of merged code is machine-authored. The daily volume has grown eightfold. In a randomized study by the nonprofit METR, experienced open-source developers using AI tools on their own repositories took about 19 percent longer to complete tasks – despite expecting the tools to help (METR, February 2026). Stanford found that agentic tasks consume wildly more tokens without producing reliably better results. And compute demand keeps behaving in ways the providers themselves describe as surprising and constrained.

The recursive loop can cut both ways – Google DeepMind got real efficiency gains from Gemini optimizing its own training kernels (Google DeepMind, 2025). The claim is not that AI makes everything less efficient every year.

THE PLAUSIBLE MECHANISM

The more defensible claim is that recursive AI development may be creating a new class of opaque technical debt: machine-written scaffolding that works, that nobody fully wants to unwind, and that may be harder to detect before it shows up in capacity planning or the bill.

Sometimes the loop optimizes. Sometimes it bloats.

The bill arrives

For the customer, the price of traditional enterprise software is largely decoupled from the engineering decisions behind it. Salesforce doesn’t charge you 10x more next month because its codebase had a bad week. That has never been true of AI – and the gap is widening.

The problem is not that AI is expensive. It is that frontier-model economics are unstable in a way that no pricing model cleanly absorbs. Subscriptions hemorrhage margin. Usage-based billing passes the volatility to the customer. Neither side can plan well, because cost is driven by behavior that is variable, opaque, and only loosely tied to output quality.

Fable was supposed to be Anthropic’s public showcase for what its best model could do. Instead it became a three-day demonstration of what its pricing model was already signaling it might not sustain. The government killed it for security reasons. The economics were already doing the job.

Frontier AI is not just expensive. It is expensive in ways that are volatile, opaque, and increasingly hard for either side to predict. If the builders are now using AI to write the systems behind the next wave of AI – and Anthropic says 80 percent of its merged code is now authored by Claude – that instability may not be a temporary scaling problem. It may be part of the stack.