A year ago the sales pitch of the AI industry, repeated in every keynote, was universal: this technology will be a blessing for all of humanity. In June 2026 the two most capable models on the planet spent significant parts of the month in a government-supervised vault, and the most capable one of all is now available to roughly a hundred vetted organisations. Something changed — not gradually, but in about four weeks — and the honest question is not whether someone is to blame but what machine produced this outcome. Because the uncomfortable answer, argued below, is that no one needed to plan it. The incentives were already pointing the same way, all of them, and June was simply the month they clicked into gear together.
Current Conditions
What Actually Happened
The timeline first, stripped of spin. On June 9, Anthropic launched Claude Fable 5 as its first generally available Mythos-class model — and, in the same breath, Mythos 5, the identical model with fewer safeguards, reserved for vetted cyber-defence and critical-infrastructure organisations. The framing from day one was explicit: this model is powerful enough to help attackers, hence the guardrails, hence the restricted twin. The guardrails promptly overshot — early users reported basic biology questions triggering safety fallbacks — while security researchers complained the opposite: the model refused legitimate defensive work. And a separate disclosure did quiet damage: for requests classified as frontier-LLM development, the system card described interventions that degrade the model’s effectiveness without notifying the user — silent handicaps in a paid product, as the community put it, affecting a stated 0.03 percent of traffic but a much larger share of trust.
Three days after launch, at 5:21 PM Eastern on June 12, Anthropic received a US export-control directive citing national-security authorities: suspend all access for any foreign national, inside or outside the United States — including Anthropic’s own foreign-national employees. The company shut both models off globally within hours. Reporting traced the trigger to a jailbreak claim from a rival company, and noted the government had previously pressed Anthropic to delay the release. Eighteen days later, on June 30, the controls were lifted; Fable returned to all users, with a temporary usage sweetener attached. Mythos remains accessible to roughly a hundred approved US organisations. In between, organisations worldwide discovered what a hard dependency feels like: automations that had auto-adopted the newest model failed silently — no fallback, no alert, just a capability that was there on Thursday and gone on Friday.
Then the part that converts an incident into a regime. In late June, OpenAI launched GPT-5.6 — Sol, Terra, Luna — not publicly, but as a limited preview to approximately twenty partner organisations whose identities were shared with the government, under an executive order asking that the most powerful models be presented for government review ahead of public release. Access, per OpenAI’s own CEO, was being approved customer by customer. OpenAI wrote, in its own launch material, that it does not believe this kind of government access process should become the long-term default — and proceeded under it anyway. Whatever June 12 was, it was not an Anthropic anomaly. It was the first enforcement of a new arrangement in which frontier models are treated as controlled goods, reviewed before release, and gated by lists.
The Lens: Bootleggers and Baptists
Regulatory economics has owned this pattern since 1983, when Bruce Yandle — then at the Federal Trade Commission — asked why regulation so often ends up serving the industries it supposedly disciplines. His answer became a classic: durable regulation arises when two groups want the same rule for opposite reasons. The baptists supply the moral case — Sunday alcohol bans protect the sabbath and the vulnerable. The bootleggers quietly profit — a ban on legal Sunday sales hands them the market every Sunday. The two never need to meet, coordinate, or even know of each other. The sermon provides the cover; the incentive does the rest; the regulation persists because both constituencies defend it.
Now look at frontier AI and notice something Yandle’s original never had: here, the baptist and the bootlegger are frequently the same organisation. The leading labs sincerely argue — and there is little reason to doubt the sincerity — that frontier models are dangerous, that oversight is needed, that uncontrolled proliferation is the nightmare scenario. Anthropic has made government-supervised safety a founding identity; its launch material for Fable said openly that the model was powerful enough to require unprecedented restriction. And the same argument, wherever it lands as policy, functions as a moat. Government review before release is a fixed cost that a two-hundred-person startup cannot pay and a hyperscaler barely notices — OpenAI casually mentioned spending 700,000 GPU-hours on jailbreak-hunting alone, a line item larger than most companies’ entire compute budget. Vetted-access programs make the incumbent labs the gatekeepers of the most valuable capability tiers. Warnings against local, uncontrolled models — sincere as safety analysis — are simultaneously arguments against the only distribution channel the incumbents do not own. The sermon and the moat are not merely compatible. They are load-bearing for each other.
No one needs to play the game for the game to be played. The sermon provides the cover, the incentives do the work, and the outcome looks designed without a designer.
The Structural Thesis — and Why It Beats the Conspiracy
Which brings us to the version of this argument worth holding, and it is deliberately not the sinister one. The weak thesis says the labs and the government choreographed this — manufacture a danger narrative, trigger a lockdown, return the product with a higher price tag and a taste for scarcity. Against that thesis stands inconvenient evidence: Anthropic lost eighteen days of revenue on its flagship at launch — the most expensive possible moment — and objected publicly; OpenAI criticised the review regime in its own launch post; both companies visibly absorbed damage. Staged plays are usually cheaper for the actors.
The strong thesis needs none of that, which is precisely why it is more disturbing. Consider the forces in play, each independently real, each independently defensible: a danger narrative that is partly true (the models genuinely are dual-use, and the same week Fable was criticised as too restrictive for defenders, it was pulled for being too capable in hostile hands); compliance and review costs that scale down badly, structurally advantaging the largest labs; datacentre demand that inflates the price of exactly the hardware a private citizen would need to run capable local models; and a policy environment in which every incident ratchets oversight upward and no incident ever ratchets it back down. None of these forces requires coordination with any other. All of them push the same direction: away from AI as an open commodity, toward AI as licensed infrastructure — tiered, metered, reviewed, and gated. Readers of Order Without a Ruler will recognise the shape inverted: order without a ruler was the optimistic case, spontaneous coordination producing freedom; this is capture without a conspirator, the same bottom-up mechanics producing concentration instead. Aligned incentives are more durable than any cartel, because a cartel can be exposed and prosecuted. An incentive gradient can only be pointed at — and pointing at it changes nothing.
Life in the Metered Lane
For the working user, the new regime arrives not as politics but as arithmetic. Fable’s API pricing runs roughly twice Opus 4.8’s, and subscribers describe a now-familiar choreography: new capability appears inside paid plans briefly — a launch window, a post-restoration goodwill period — then migrates toward higher tiers and tighter limits. The rational response has already become a skill in itself: model routing. Serious users now split their workflows — routine scaffolding to a cheap fast model, standard work to the mid-tier, and the frontier model reserved, at carefully chosen effort levels, for the problems that actually need it. It works. It also means the headline experience of AI in 2026 is no longer “what can it do” but “what can I afford to ask it” — a sentence that would have sounded absurd in the blessing-for-all-mankind era, roughly one year ago.
And the exit routes bend the same direction. Local models are the honest escape hatch, and honesty requires sizing it correctly: a used high-VRAM GPU around 750 euros runs genuinely capable mid-size open models — that door is real and open. But anything approaching frontier quality demands hardware measured in five figures, and the datacentre buildout of the closed-model providers keeps repricing exactly those components upward. Meanwhile the labs’ safety position — sincere, published, consistent — holds that uncontrolled local models are the very risk their oversight exists to prevent. Follow the loop: the danger narrative justifies the gates; the gates create the tiers; the tiers push users toward exits; the exits are priced up by the gatekeepers’ own infrastructure demand and argued down by the gatekeepers’ own safety doctrine. No single link in that chain is illegitimate. The chain as a whole has a direction.
One disclosure, because this blog has spent whole pieces arguing that verification must not be outsourced: this analysis was researched and drafted with Claude Fable 5 — the model that spent eighteen June days in the vault, built by a company whose incentives this piece dissects. The reader should weigh that exactly as seriously as it sounds, and check the load-bearing facts independently. The timeline is public record; the lens is forty years old; the conclusion is yours to draw.
What to Actually Take From This
The debate is stuck between two useless poles — “necessary safety” and “evil cartel” — while the durable takeaways sit in the structure, not the intent.
The kill switch is no longer hypothetical. The One Machine argued that who can switch your system off matters more than how fast it runs. June was the live demonstration: the most capable available model, removed globally within hours, by letter. Any workflow with a single-provider frontier dependency now has a documented failure mode — and no excuse for not having a fallback.
Distrust alignment of sermon and moat — without inventing a cartel. The safety arguments are partly true; the dual-use capability is real; the labs demonstrably absorbed costs and objected. And still every force in play — narrative, compliance cost, hardware price, review regime — pushes toward incumbent-gated AI. You do not need bad actors for a bad equilibrium. That is precisely what makes it hard to reverse.
Position for the licensed frontier, hope for the open one. Practically: redundancy across providers, model routing as a budgeting skill, a right-sized local fallback, and eyes on the ratchet — oversight that only ever tightens is a trend, not an event. The blessing-for-all framing is a year old and already reads like another era. Plan for the era that actually arrived.
Instrument Check — Worth Your Attention
Read — “Bootleggers and Baptists,” Bruce Yandle (Regulation, 1983). Seven pages from a former FTC economist explaining why regulation so often serves the regulated. The moral-cover-plus-quiet-beneficiary mechanism has explained forty years of policy since, and it maps onto AI oversight with almost embarrassing precision — with the twist that in AI, both roles are often played from the same building.
Study — the June 2026 primary record, both labs’ own statements. Read Anthropic’s suspension and restoration statements and OpenAI’s GPT-5.6 preview post side by side. Two competitors, weeks apart, each objecting to the review regime in writing while operating under it. The convergence of language is the single strongest piece of evidence that this is a regime, not an incident — and it is all on the record.
Follow — the choke point and the check: The One Machine and The Outsourced Check. The first argued that choke points are swords that double as leashes, and that “who can switch you off” beats “how fast does it run” — before June supplied the live demo. The second is the discipline this piece leans on: separating what is verified from what is merely repeated, especially when the source describing the vault is the model that was inside it.
Flight Log — Dispatch From Altitude
Aviation has been running this experiment for seventy years, and the results are worth reading carefully, because they refuse to flatter either side of the argument. Ask why only two companies on Earth build large airliners and the popular answer is engineering difficulty. The real answer is certification. Bringing a new large transport aircraft to market means a decade and several billion dollars of regulatory proof before the first ticket is sold — tens of thousands of pages, thousands of test conditions, every system demonstrated against failure modes most engineers will never see in a career. The safety case for every one of those requirements is written in blood; almost none of them is arbitrary. And their combined weight is the most effective competition barrier ever constructed. Nobody designed it as one. It emerged.
Notice the structure, because it is the bootlegger-baptist machine with rivets. The baptists are real: regulators, investigators, the accumulated dead of a century of accidents — the moral case for every rule is sincere and mostly correct. The bootleggers never had to lobby for the outcome: the incumbents comply, at enormous cost, and that cost is precisely what no newcomer can survive. Startups have tried — with money, with talent, with better ideas — and died not in the wind tunnel but in the paperwork. The sermon is true, the moat is real, and they are the same document. No conspiracy, no cartel, no meeting where anyone decided that safety regulation should also be a fortress. Aligned incentives, compounding for decades, produced what no cartel could have sustained: a duopoly defended by the most legitimate argument in the world.
And here is the part both sides of the AI debate should sit with: it worked. Flying became the safest form of travel in human history because of that regime, not despite it — and the price was a market where innovation slowed, where two suppliers set the terms, and where the certified lane is the only lane. Aviation quietly accepted a trade nobody ever voted on: radical safety in exchange for radical concentration. Perhaps it was the right trade for aluminium tubes full of people at forty thousand feet. Whether it is the right trade for intelligence itself — for the tool every other tool now runs on — is the question June 2026 put on the table, and it deserves better than to be answered by default, by drift, by the silent compounding of everyone’s locally reasonable incentives.
Because that is how the licensed frontier actually gets built — not by decree but by accretion, one justified requirement at a time, each defensible, none reversible, until one day the notion of building outside the system sounds not illegal but simply unserious. Aviation crossed that line so long ago that nobody remembers it as a line. AI crossed something in June, and the crossing is still visible — which means it can still be discussed, priced, and, where the trade is bad, refused. The cockpit teaches respect for regulation written in blood. It also teaches that you check what every system is actually doing, no matter how good its intentions — because the autopilot of incentives flies very smoothly, right up until you notice where it has been taking you.