The AI Caste System: Restrictive Models and Cognitive Stratification

On June 1, 2026, GitHub quietly ended one of the most important subsidies in the history of software development. The flat-rate subscription model — pay your $10 or $39 a month and use it as freely as you like — was replaced with token-based billing. Every input, every output, every cached context window now runs a meter.

The reaction from the developer community was immediate and overwhelming. The GitHub community discussion thread drew 904 downvotes against 22 upvotes in the first 72 hours, one of the most lopsided responses in the forum's history. Screenshots circulated on Reddit and X showing projected monthly bills jumping from $29 to $750, from $50 to $3,000, and in some extreme agentic coding workflows, even higher. TechCrunch called it the end of “the golden age” of GitHub Copilot.

The people hit hardest were not experienced engineers — those who primarily use Copilot for inline autocomplete saw almost no change, since code completions remain unlimited. The casualties were the vibe coders: the non-developers who had built entire workflows around Copilot's agent mode and chat interface, describing what they wanted in plain language and letting the AI assemble it. For that group, agentic sessions consuming thousands of tokens per interaction became ruinously expensive overnight. One developer on the $39 Pro+ plan reported burning through 8% of their monthly credit allotment in just two hours.

GitHub's own announcement, written by Chief Product Officer Mario Rodriguez, was honest about why: the unlimited model had become untenable. The subsidy era, as I predicted, is over.

Some people pushed back on this thesis, arguing that cheaper alternatives exist: open-source models, local inference, budget providers like DeepSeek and Mistral. That argument is not wrong, but it is incomplete. The cheaper and open models have improved — sometimes dramatically — but they have not converged with frontier capability, and the gap on the tasks that matter most (long-context reasoning, advanced agentic behavior, deep coding assistance) remains wide.

The first major signal came from Anthropic. In April 2026, the company introduced Claude Mythos — a model they described as occupying an entirely new capability tier. Anthropic chose not to release it broadly, deploying it instead through a limited program called Project Glasswing, available only to a small number of vetted, high-value organizations.

Then, on June 9, 2026, Anthropic released Fable 5 — a consumer-facing model built on the Mythos architecture but equipped with safeguards designed to prevent access to its most dangerous capabilities. Three days later, the U.S. government ordered Anthropic to pull it entirely. The Commerce Department issued an export control directive suspending all access to both Fable 5 and Mythos 5 by any foreign national — not just people outside the United States, but any non-citizen anywhere on Earth, including Anthropic's own foreign national employees.

Over 100 cybersecurity executives, CISOs, and researchers signed an open letter to Commerce Secretary Howard Lutnick arguing that the ban had “taken the best models away from defenders” without justifiable security grounds — while China continued shipping open-weight frontier models without equivalent restrictions. As of this writing, two weeks after the directive, Fable 5 and Mythos 5 remain offline for all users worldwide.

Then, almost simultaneously, came the release of GPT-5.6. The Trump administration asked OpenAI to restrict the rollout to a small group of trusted partners — approximately 20 companies whose participation had been reviewed and approved by the government. This marks the first time in U.S. history that the government has preemptively asked an American AI company to restrict a model's public release before launch.

OpenAI's own statement made clear the company was uncomfortable with the arrangement: “We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” But they complied anyway.

At the top sits a narrow priesthood: government agencies, the Pentagon, a handful of vetted defense contractors and cybersecurity firms, and the small number of elite technology companies that have secured the right relationships and contracts. These organizations are working with AI systems that the rest of the world is not allowed to see, let alone use.

We are, in effect, allowing a small number of entities — some government, some private, all extremely well-resourced — to develop a structural lead in cognitive capability that may become very difficult to reverse. The historical analogies are not encouraging. When literacy was concentrated, it took centuries and a technological revolution to democratize it.

For most of human history, literacy was not universal. The ability to read and write was concentrated in a small class of specialists — clergymen, scribes, scholars — who mediated between the general population and the texts that organized society. These specialists had access to knowledge, law, theology, and medicine that the ordinary person could not access independently.

The printing press disrupted that hierarchy. It did not eliminate it immediately — for generations, elite institutions controlled what was printed and distributed. But the direction of travel was unmistakable: knowledge was decentralizing.

AI is doing something similar to cognition itself. Just a few years ago, the ability to draft a complex legal brief, write production-quality code, diagnose a rare medical condition, or architect a building required years of specialized education and in most cases a substantial fee to access through a professional. Today, a capable AI can do all of these things in seconds.

This is not magic, but it looks like magic to most people — in precisely the same way that a monk's ability to read a Latin manuscript looked like magic to an illiterate peasant in the 12th century. And now, as with the manuscripts of old, those outputs — specifically the most powerful versions of them — are being concentrated in the hands of a small number of institutions who have the resources, relationships, and government clearance to access them.

We are building a new priest class. It is made up of defense contractors, government agencies, and elite technology companies rather than clerics. The magical books are language models with cybersecurity and scientific reasoning capabilities that exceed anything previously available to the public. And the rest of us are being told, politely but firmly, that these books are too powerful to let us read.

The standard arguments for restricting powerful AI center on safety. The models are genuinely capable of helping identify software vulnerabilities, potentially accelerating the development of dangerous biological or chemical agents, and assisting with a range of activities that could cause serious harm at scale. These are not imaginary concerns.

But the argument cuts both ways. Restricting these tools does not eliminate the underlying capabilities from the world — it simply determines who has access to them. If a government-vetted contractor can use Mythos to find software vulnerabilities, and a foreign adversary's intelligence service is using its own equivalent model to find the same vulnerabilities, the only entity left without access is the independent security researcher, the small startup, and the ordinary developer who might otherwise have used the tool defensively.

The ban on Fable 5 and Mythos 5 did not make the world safer for ordinary people. It made it safer for the institutions at the top of the caste system, who retain access to the frontier, while leaving everyone else slightly more exposed than they were before.

There is also a more structural danger that gets less attention: concentration of decision-making power. When a small number of institutions hold exclusive access to the most capable AI systems, those institutions gain asymmetric advantages in every domain those systems touch — legal research, medical diagnosis, scientific discovery, financial modeling, policy analysis, intelligence gathering. The organizations that can run the most powerful models fastest will consistently out-think, out-plan, and out-maneuver those that cannot.

The fourth warning is the one that hits closest to the ground: the death of the level playing field in business. Before AI, a software firm with five employees could compete with one that had five hundred, if the smaller firm had the better engineers. That equalizer is now being dismantled. A firm with a direct contract to deploy Mythos-class models is not a little bit faster than the boutique firm running on capped consumer tiers. It is operating in a categorically different cognitive environment.

There is, however, a countervailing force worth naming. If AI contracts become expensive enough that their cost exceeds what a traditional engineering team would have cost, the economics of AI adoption start to invert. The firm that chose not to buy in — that kept its engineers, maintained its craft, and built products where every line of code was written and understood by a human — may find itself with something increasingly rare and valuable.

I call this the Human Only House. As AI-generated code proliferates, the software product that can demonstrate it was built entirely by human hands will carry a premium that no AI-assisted competitor can match — not because it is more feature-rich, but because it is more trustworthy. Because the people behind it actually understand what is inside it. Because it is not subject to what I think of as the Blinded Mind Problem: the growing reality that engineers who rely entirely on AI-generated code are building structures they cannot fully read, debug, or defend.

I am not arguing that frontier AI should be entirely unrestricted. The concerns about genuinely dangerous capabilities are real, and some form of oversight is probably necessary as these models become more powerful. But oversight is not the same as exclusive access. There is a meaningful difference between saying “these models require safety review before broad release” and saying “these models will only ever be available to twenty government-approved companies.”

The current trajectory is sliding toward the latter without an honest public conversation about whether that is what we want. The gatekeepers of technology have always been worth watching. In every era, the people who control the most powerful tools of their time also shape the social and political order that surrounds those tools. The choices being made right now — about who gets access to the frontier, on what terms, and subject to what oversight — will shape what kind of society we build with this technology.