{"slug": "the-ai-dark-age-will-run-on-a-government-switch", "title": "The AI Dark Age Will Run on a Government Switch", "summary": "The US government has established a regime controlling the release of frontier AI models, exemplified by pulling Anthropic's Claude Fable 5 offline on June 12, restricting OpenAI's GPT-5.6 to government-approved partners, and clearing Anthropic's Mythos 5 only for vetted US institutions. This marks a shift from a grudge against one company to a system where the government holds the switch on the most capable AI models.", "body_md": "# The AI Dark Age Will Run on a Government Switch\n\nWhen the US government [pulled Claude Fable 5 offline](https://www.anthropic.com/news/fable-mythos-access) on June 12, I told myself a story that made it smaller.\n\n*Anthropic*had refused the Pentagon back in February, this was the payback, narrow and personal, the kind of thing that lands on one company and stops there. I more or less believed it.\n\n**Two weeks later the story fell apart.**\n\nOn Friday, June 26, the White House asked *OpenAI* to do the gentler version of the same thing: hold its new flagship, *GPT-5.6*, back from the public and ship it only to a small group of government-approved “trusted partners.” *OpenAI* agreed. That same Friday, the Commerce Department cleared *Anthropic*’s strongest model, *Mythos 5*, for release to **about 100 vetted US institutions** on a list the Commerce Secretary says he can change [“at any time”](https://thehill.com/policy/technology/5943549-anthropic-mythos-5-access/). *Fable 5*, the model hundreds of millions of people were actually using, **stayed dark.** As of June 27 it still is, with **no date attached.**\n\nSo inside a single month: the best model from the lab that defied the Pentagon, gone for everyone. The best model from the lab that cooperated, parceled out to a government list. And the newest model from a third lab, held back at the government’s request before the rest of the world could touch it. **That is not a grudge against one company. That is a regime.** The outlet that obtained the *Mythos* letter called it exactly that: [the beginnings of a system that gives the US government control over the release of frontier AI models](https://www.semafor.com/article/06/27/2026/us-releases-powerful-anthropic-model-mythos-to-some-us-companies).\n\nI should say where I stand, because it changes how you read the rest. I build AI agents for a living and I use these models every single day, *Claude Code*, the frontier stuff, all of it. I am about as pro-AI as a person gets, and a lot of what these labs have shipped is the best software I have ever touched. **None of what follows is doom about the models.** **The models are great.** The thing worth worrying about is **who now holds the switch behind them.**\n\n## The Two Weeks That Turned a Grudge Into a Regime\n\nThe “lawfare against *Anthropic*” reading was the easy one, and I was reading it that way too. If the only thing happening were punishment for the lab that said no, *OpenAI* would have been left alone. *OpenAI* took the Pentagon deal. *OpenAI* is the cooperative one. And *OpenAI* got [asked to gate GPT-5.6 anyway](https://openai.com/index/previewing-gpt-5-6-sol/), under the same cybersecurity rationale, inside two weeks.\n\n**The Pentagon fight explains how hard** The gate exists for everyone.\n\n*Anthropic*got hit, not whether the gate exists.The mechanics differ, and the softer version is the more revealing one. *Anthropic* got a binding Commerce export-control order and had to switch *Fable* and *Mythos* off worldwide. *OpenAI* got a request from the White House and said yes: no order, no recall, just an agreement to **keep its best model away from the public until the government signs off on who gets in.** *OpenAI* did not even pretend to like it. In its own announcement it said it does not believe this kind of government access process should [“become the long-term default,”](https://openai.com/index/previewing-gpt-5-6-sol/) because it “keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” **They are right, and they did it anyway.** That is what a norm looks like once it sets. Even the company that hates it complies, because the alternative is worse.\n\nAnd these are not toys we are talking about. By *Anthropic*’s own published benchmarks, *Fable 5* and *Mythos 5* were the most capable models it had ever shipped, **ahead of GPT-5.5 and even Anthropic’s own Opus 4.8** on the hard coding and offensive-cyber tests. The best tools for real work that exist now route through one government’s approval, and that approval has already been switched off, granted, and withheld more than once in thirty days.\n\n## A License by Another Name\n\nThree weeks before any of this, [on June 2 the President signed an executive order](https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/) that went out of its way to promise the opposite. In plain text, it says nothing in it authorizes “a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models.” **Voluntary cooperation only. No licenses.**\n\nThen watch what voluntary turned into. The day before that order, the Commerce Secretary had already put *Fable* and *Mythos* under export controls by letter. By June 26 the same Secretary was handing out access to *Mythos* against a government-held whitelist, while reserving in writing [“the right to reevaluate and adjust the scope of license requirements”](https://thehill.com/policy/technology/5943549-anthropic-mythos-5-access/) whenever he decides circumstances have changed. A model can ship to this company and not that one. The list can grow or shrink at one official’s discretion. **A system where the government decides, name by name, who is allowed to run a piece of software is a license, whatever the order calls it.** The EO swore off licensing on June 2 and the government was exercising one by June 26.\n\nAnd sit with what that discretion actually is. When a powerful new model clears the bar, the same approval that lets it out can route it to one company and leave that company’s direct competitors waiting, on no published criteria, decided by the same people the lobbyists are paid to reach. **The government now holds a lever that hands a chosen firm a head start and makes its rivals come asking for one.** Preferential treatment was dangerous long before AI, and Washington runs on access and favors. Point a lever like that at the technology every company is now betting on, and there is no version of this where it stays neutral for long.\n\nThe vocabulary is already in place. “Covered frontier models.” Trusted-partner lists. Thirty-day windows where a model goes to the government before it goes to you. **We are watching the permission layer for AI get built in real time, by letter, on the fly,** and the labs are learning to ask before they ship.\n\n## We Built the Switch Ourselves\n\nEvery imagined version of an AI catastrophe is a story about escape. The model gets loose, breaks containment, slips the lab. It is a comforting shape, because it puts a wall between us and the danger and lets us picture the danger having to climb over it.\n\nThere is no wall. We are wiring these models into everything ourselves, willingly, one convenient tool at a time. *Claude* ends up in a hospital’s records because it saves a nurse twenty minutes, and the day it saves that twenty minutes it is in for good. The same thing is happening in banks, law firms, IDEs, customer support queues, and government offices. Every one of those integrations is a small, sensible decision on its own. The sum is one company’s product threaded through the machinery of an entire economy, and **the thread does not pull back out.**\n\nWhich is exactly why the switch matters. The risk was never that the model would get out. It is that we keep inviting it in, deeper every quarter, and we have just learned that the off-switch behind it sits on a desk in Washington. **We spent three years handing our most important workflows to tools that one government can now turn off with a letter.** Nothing has to break containment when we keep opening the door from the inside.\n\n## Cut Off by Your Passport, with Nothing to Fall Back On\n\nI am Belgian. Under the *Fable* order that makes me a “foreign national,” which put the best coding model on Earth off-limits to me because of my passport, not because of anything I ever did with it. When it happened I [wrote that the precedent would outlast whatever the jailbreak was](https://eigenwise.io/writing/the-jailbreak-in-every-model). Two weeks of precedent later, here we are.\n\nThe passport part is the small, petty annoyance. The real weight of it shows up when you ask the next question: when Washington controls the supply of the best models, what does everyone outside the approved list fall back on? For Europe, where I live and work, the honest answer is **nothing in the same tier.**\n\n*Mistral* is the best we have and it is genuinely good. But you cannot do heavy agentic coding with it and expect *Fable* or *Opus* or *GPT-5.x* results, and I say that as someone who wanted it to be true and tested it on real work. The reason is boring and structural: **budget and compute.** *Mistral*, the most valuable AI lab in Europe, [raised €1.7 billion](https://mistral.ai/news/mistral-ai-raises-1-7-b-to-accelerate-technological-progress-with-ai/) in its 2025 round. **The American leaders raise multiples of that in a single deal and spend it on compute several times over every year.** Europe is finally pledging real money, hundreds of billions on paper for infrastructure and “gigafactories,” but that is data centers years out, not a model you can run on Monday when the American one goes dark. The sovereign fallback everyone assumes is sitting there, ready, simply isn’t.\n\nThe other answer people reach for is Chinese open weights. *GLM*, *DeepSeek*, *Qwen*. Genuinely impressive, much cheaper, and because the weights are open you can run them yourself with no one positioned to switch them off, which is the real argument in their favor. For heavy use, though, they don’t close the gap. I gave *GLM 5.2* a serious run on the promise that it matches *Opus* at a fraction of the cost, and I hit my usage limits about **twice as fast**, because the model burns more tokens to reach the same answer. **Cheaper per token isn’t cheaper per task,** and the difference bites exactly where a power user lives, at the limit. And that open hatch is not guaranteed to stay open. If Beijing ever decides its frontier models are strategic assets the way Washington just did, the gate closes from the other side too, and then it is two superpowers rationing their best models while everyone else negotiates for access to tools they did not build and cannot replace.\n\n## What the Gate Does to the Rest of Us\n\nI don’t want this read as an anti-*Anthropic* piece. *Anthropic* is the lab that said no, it refused two Pentagon red lines the others folded on, and *Claude* is most days the most capable and most decent tool in its class. Loving the model and trusting the government now sitting upstream of it are two different bets. The worry was never the company. **It’s what one government can now do with a product hundreds of millions of people depend on, and what that does to everyone downstream of the decision.**\n\nStart with instability. Banks, hospitals, courts, and half the software industry are now building on models that can be switched off by letter, with no notice and no appeal. **A foundation that can be revoked with a phone call is not a foundation,** and we poured the economy onto it anyway. In its first month this regime has already switched one model off, handed another to a list, and held a third back. That is the ground everyone is now standing on.\n\nThen the inequality, which is the part that should actually worry you. The obvious version runs between countries: Americans and a few allies on the inside, Europe and most of the world cut off by passport with nothing of their own to fall back on. The quieter version runs **straight through the middle of America itself.** The list is about a hundred institutions. If you are a US startup, a solo developer, a small clinic, a school, a shop, any of the millions of businesses not on it, you are outside the frontier too, sorted there by a government decision instead of a price. **And AI advantage compounds.** The handful of players holding the best models pull further ahead every quarter, and the distance behind them hardens into a wall. We are about to learn what it does to a society when the most powerful tool of the decade goes to a hundred approved names and gets rationed to everyone else.\n\nThat is what a dark age actually is: **knowledge that existed in the open, now rationed, and a society reorganizing itself around who got a copy and who did not.** The lights do not have to go out anywhere for it to happen. The technology will keep getting better, and eventually it gets out, because it always does. The damage gets done in the years before it does, **in the instability and the inequality we are pouring the foundations for right now,** and none of that is going to show up on a benchmark.\n\nOne reason I am not all doom about this: **most real-world AI value was never going to come from having the single best model anyway.** It comes from the setup around it: the context you give it, the guardrails, the fit to your actual problem, the unglamorous part no leaderboard measures. The frontier getting rationed genuinely matters, but for the large majority of what companies need AI to actually do, a model you can get today, wired up well, clears the bar.\n\nThat part is what I work on. At [eigenwise.io](https://eigenwise.io) I help companies get real value out of their AI stack, the setup work that turns an accessible model into something that pays for itself. If that is the problem you are sitting with, that is where to find me.", "url": "https://wpnews.pro/news/the-ai-dark-age-will-run-on-a-government-switch", "canonical_source": "https://eigenwise.io/writing/the-ai-dark-age-government-switch", "published_at": "2026-06-27 00:00:00+00:00", "updated_at": "2026-06-27 17:40:12.835083+00:00", "lang": "en", "topics": ["ai-policy", "ai-safety", "large-language-models", "ai-ethics"], "entities": ["Anthropic", "OpenAI", "White House", "Commerce Department", "Claude Fable 5", "GPT-5.6", "Mythos 5", "Pentagon"], "alternates": {"html": "https://wpnews.pro/news/the-ai-dark-age-will-run-on-a-government-switch", "markdown": "https://wpnews.pro/news/the-ai-dark-age-will-run-on-a-government-switch.md", "text": "https://wpnews.pro/news/the-ai-dark-age-will-run-on-a-government-switch.txt", "jsonld": "https://wpnews.pro/news/the-ai-dark-age-will-run-on-a-government-switch.jsonld"}}