The human impulse to ration access to power has been a recurring pattern throughout history. In medieval Europe, the Catholic Church kept scripture in Latin, a language the laity could not read, so that interpretation, and thus authority, remained the clergy's monopoly. The printing press and vernacular Bibles shattered that arrangement. We are now constructing a comparable enclosure around artificial intelligence, and we call it "Limited Release": a deliberate stratification of cognitive capability determined by who can pay the fees, and, increasingly, by who the state has cleared to receive it.
Consider the sequence of events. On June 9, 2026, Anthropic released Claude Fable 5 and Mythos 5, announcing the model "exceeds those of any model we've ever made generally available," with the lead widening "the longer and more complex the task". Three days later it was gone. On June 12, the Trump administration issued an export-control directive suspending access for "any foreign national," forcing Anthropic to disable both models for every customer worldwide on national-security grounds. The frontier was handed to the public and clawed back in seventy-two hours. The most capable tier, Mythos 5, had never been broadly available at all. It was distributed through a restricted trusted-access program called Project Glasswing, a coalition that hands an unreleased frontier model directly to AWS, Apple, Cisco, Google, Microsoft, JPMorganChase, and other enterprises deemed defenders, with the government consulted on its cyber capabilities. GPT-5.6 has now internalized the lesson and skipped the misstep entirely: it "launched for OpenAI enterprise partners for testing" on June 25, with the public release deferred to the second week of July. This is a "limited release" by design. The frontier is not released to humanity. It is released to the procurement departments of enterprises, by government dictate. The citizen waits in the antechamber until Big Brother decides the people are ready.
The arrangement is a clearance regime sorting humanity by permission. ChatGPT Enterprise carries a 150-seat minimum and a roughly $108,000 annual floor, a threshold no individual crosses. Fable 5, when it briefly existed, ran $50 per million output tokens—5.5 times Gemini 3.5 Flash. This is a price only an institution amortizes painlessly. The individual, if granted access at all, gets the throttled tier; GPT-5.6 may even be "less token efficient," meaning ordinary users burn through quotas faster on the same task. The Glasswing member gets autonomous agents that find thousands of zero-days no human could; everyone else gets the dulled, delayed, state-sanctioned remainder. This is not a free market distributing tools by merit. It is a structure that hands the sharpest cognitive instruments to those already capitalized and politically approved, while rationing degraded versions to the rest. This is the precise definition of a caste.
The historical parallel sharpens further. Just as guilds in early modern Europe restricted technique to protect incumbents, today's labs gate capability behind enterprise contracts while the marginal worker, the solo founder, and the student receive the hand-me-downs. Advantage accrues to advantage. The justification is always the same: safety, security, the public good. This is the exact same language the Church used to keep the Word in a dead tongue.
This is why the looming proposals to outlaw or license local models should alarm anyone who values an open mind. Open-weight systems are the one escape hatch from the enterprise antechamber: the only frontier-class capability a person can run, own, and inspect without a procurement contract or a regulator's blessing. And they are improving fast. To criminalize self-hosted intelligence under the banner of safety would be to burn the printing press to protect the scriptorium and ultimately severing the people's last route to the frontier precisely as that route becomes viable.
There is, however, an dangerous alternative at the gate. Chinese open-source models now constitute roughly 30% of global AI usage, led by Alibaba's Qwen and DeepSeek. Qwen crossed one billion cumulative Hugging Face downloads by March 2026, faster than any model family in history. DeepSeek V4 still trails GPT-5.6 reasoning by three to six months and scores 52 on the Artificial Analysis Intelligence Index against the frontier's higher marks. But it delivers output roughly 35 times cheaper than GPT-5.5, and it does so as weights you can download tonight. The vernacular Bible, it turns out, may be printed in Hangzhou.
The uncomfortable conclusion: the West is building the cathedral while China prints the pamphlets. If frontier capability remains a subscription enjoyed by the few, and if local models are legislated out of existence to keep it that way, the democratizing force will not come from Silicon Valley. It will come from whoever gives the people a model they can actually own.
