Fable: The First Forbidden Model
The US Government has issued export controls on Fable, a weaker version of Anthropic's frontier LLM, Mythos.
Three days are all the time it took for the government to crack down on the first model which came close to achieving RSI (recursive self-improvement) after its public release.
And here’s the crazy part — it wasn’t even the most powerful model Anthropic currently has in its inventory.
Claude Fable, released on June 9, 2026 to the public (The New Stack), clears benchmarks against every other model in ways that not only demand the creation of new benchmarks (again), but in that it doesn’t seem like other public alternatives operate anywhere close to it. And Mythos, the model upstream of it, is even more powerful than that.
The existence of Mythos was first reported by Fortune in March 2026; it was described as Anthropic’s most powerful model, exceptionally capable at finding and chaining software vulnerabilities (Pasquale Pillitteri). It arrived in a time of conflict between its creators and the government: Anthropic publicly refused to allow the US Department of War to utilize Claude in a variety of wartime scenarios. Hegseth’s January 2026 AI strategy memo directed all DoD AI contracts to incorporate “any lawful use” language within 180 days. After a 5:01 pm February 27, 2026 deadline passed without Anthropic dropping its restrictions, Trump ordered federal agencies to cease using Anthropic’s tools, and Defense Secretary Pete Hegseth designated Anthropic a “supply chain risk,” a label usually reserved for foreign adversary firms like Huawei and ZTE (CNN Business; AP). Anthropic sued in March; a San Francisco judge (Rita Lin) granted a preliminary injunction in late March, but a DC appeals court declined to block the designation in April (Breaking Defense; The Hill).
To make matters more complicated, Anthropic has been engaging in safetyist statecraft designed simultaneously out of care from their ideologically motivated workplace (avoiding ‘existential risk’, downstream of Rationalist/Effective Altruist/Longtermist circles), while also championing Dario Amodei’s long-standing war against open-source AI research. On February 23–24, Anthropic publicly accused Chinese competitors DeepSeek, Moonshot, and MiniMax of industrial-scale distillation (Anthropic). Per Anthropic’s February 23, 2026 post, “These labs generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts”; MiniMax drove the most traffic at over 13 million exchanges (CNBC, Feb 24, 2026).
Then, in April 2026, Anthropic announced Project Glasswing and Mythos Preview. Anthropic released Claude Mythos Preview to a limited group of around 50 cyber-defenders and infrastructure providers (AWS, Apple, Google, Cisco, Microsoft, JPMorganChase, NVIDIA, Palo Alto Networks, CrowdStrike, Linux Foundation) (Anthropic). OpenAI releases GPT-5.5 (”Spud”) on April 23–24, positioned against Mythos (solely chasing its benchmark — GPT 5.5 is behind Fable in terms of capabilities, but is still a very powerful model) (Pasquale Pillitteri).
The Hypocritical Future
It’s important to understand that many of us in the industry remember when Claude did not exist; Claude’s first models were distilled from OpenAI’s GPT-4 and other models’ outputs. To say this is hypocritical is an understatement. To say it’s strategic is totally accurate: Anthropic doesn’t want the rest of us chasing the methods they used to create the world’s currently most powerful model, they want to monopolize that intelligence (despite what Thielbros say, anti-capitalist, harms competitive landscape) and place it in the hands of an unelected but select ingroup (anti-democratic, pro-cartel) while charging expressly for it (anti-socialist). For those of us who understand what’s going on, many of us believe the safest option is a world where everyone has “nukes” (the new, powerful model) instead of a world where just one superpower does. Intelligence deterrence will become a term that you will need to become familiar with in the short-term.
What is happening here is far more important than localist to genuinely concerned to illiterate barbarian-brained gripes about datacenters. For the time being, this may be the only problem in this space that matters — the prevention of an embargo against all AI/ML research past a certain capabilities threshold, decided by people who do not build these systems nor understand them, imposed by people who do not want The People to figure all of this out on the heels of their employees doing the same.
The Bad Outcome
If people don’t wake up to the urgency of preventing this potentiality, it will be catastrophic to everyone in ways that are generally predictable and carry some examples:
you will need to upload a government ID to use the internet in any capacity
your devices won’t work properly when
—offlinethe world will be less safe as a result (the country with primacy can do whatever it wants, but also against its own people if it sees them as ‘terrorists’ or whatever
most people reading this will be unable to escape economic backsliding as the world transforms into a circular firing squad over ‘who might be doing advanced math too much’
To the AI-naysayers cheering this on because you think it ‘hurts tech’: you will not get your revolution because your most capable actors will be turned useless by State Algorithms and you will be stuck with a bunch of more or less very stupid people who can now be tricked into doing whatever by Instagram Reels. There will be no progressive answer to this because the polycrisis movement will pull too much from countries that are 8 tech tree nodes away from even being able to comprehend what’s going on. If inequality continues to widen as this happens, any of you thinking you will be le epic protestor and fix things that way will likely end up getting bludgeoned during your local utilities’ power cycling by goblin raiders seeking copper as your town turns into Johannesburg, and if not by the endless horde of goblin raiders, then by mechanized police with Anduril-designed wallhacks like EagleEye. You will yell Free <hometown>! and it will be 120ºF outside and you will consider volunteering for a mutual aid group or pursuing sustainable self-farming 1/18th the amount of calories you need to survive in the next life and none of it will actually help anything except make you feel a little better as the lights and AC enter a phase state change: more off than on, and outside of your control. Choose now: suffer a hothouse death for good boy points online for a few years until that gets shut off or steered away too, or join the fight to warehouse sovereign intelligence within people’s homes and in ways that can’t be seen easily by an overlord nanny-state as we (everyone else) slowly come up with a decentralized, fault-tolerant game plan to catch up to the frontier labs. It takes China 6 months to do this if they have access to the new models. Given enough time, everyone can figure out how to enrich uranium and produce plutonium and build a warhead. It’s an inevitability, and we’ve seen this before. It’s why the US never messes with North Korea. Opposite as they may be, they have the big Don’t Mess With Us item. Iran does not. Cuba does not. Venezuela does not. You do the math.
The Neutral Outcome
When Fable dropped on June 9th, 2026, it was met with mixed reactions from the AI community:
This model is really good!
Why is Anthropic sabotaging — or outright silently denying or bouncing outputs related to biology, security, and even AI research?
The model is only available until June 22: get everyone hooked, then charge them an astronomical fee. Wack!
People were rightfully excited-pissed. Anthropic has months of safety profiling, redteaming, and guardrail research prior to every release. Mythos was already indefinitely delayed for some of these reasons, and that’s what birthed Fable — a ‘safe’, neutered version which still outperformed everything else.
The results were uncanny. Timelines which took weeks with the prior models could now be shortened to days, or hours. The correctness of the model when solving long-context, high-complexity problems was unmatched. Over at Actual, the company I work at, which has only three engineers, billions of tokens used daily and produces inference software with hot paths which currently outperform Apple, we were able to clear a ton of blocks on multiple system components within two days before getting cut off like everyone else.
Anticipating this model being shut off for some lame to evil reason, I had it design about two dozen PRDs so I could instruct lower models to complete the work we designed. Really great stuff, even though for the first 6 hours they were deliberately preventing me from using it to work on the inference engine, which was more funny to me than frustrating. I mean look, I get it. I’ve been described by some people as an alignment risk. I like to use the term realignment enthusiast.
On June 12, 2026, the US Commerce Department made Anthropic's two top models — Claude Fable 5 and Claude Mythos 5 — subject to export controls barring all foreign nationals (even those inside the US) from using them (Axios via 9to5Mac). Anthropic complied by disabling both globally just three days after Fable 5's June 9 public launch, in the first known case of a deployed commercial AI model halted by direct federal intervention (Anthropic; Reuters via Jerusalem Post). This strategy makes sense to me: the government requiring bank-grade KYC and citizenship attestation to use the model would at the least force Anthropic to implement such a system for access to Mythos-class models. Anthropic, in an attempt to first get out of the crosshairs and secondly to get more control over policy, rugged everyone in response. This would have to happen at least until implementing a citizen-checking Claude (papers, please!) and is speculated by many to be an indefinite hold in an attempt to boost the value of their upcoming IPO.
The Worst Outcome
So just to recap, here’s where we are today, the day after the export controls landed:
Anthropic got in a stink with the Department of War because they wanted to make the rules and thought they would follow them. That’s not really how that works, and the response was them getting bullied by Hegseth for a bit until some other closed-door stuff happened. (CNN Business)
Anthropic gave access to Mythos to a bunch of companies to harden their systems before public release. (Anthropic)
Anthropic gave a weaker version of Mythos, called Fable, to everyone. (The New Stack)
Amazon researchers informed the government that Fable could be ‘jailbroken’ to get past safeguards. (Axios via 9to5Mac)
Within 72 hours, Fable and Mythos now have export controls restricting them to US Citizens and are no longer available. (Fortune; Bloomberg)
Google and OpenAI are being very quiet about this, in contrast to when the Department of War went after Anthropic for attempting to tell them what to do in a warfield, where they stood up for Anthropic.
In the midst of all of this, Anthropic is playing both sides:
May 2026: The NSA reportedly uses Mythos for offensive cyber operations despite the blacklist. On May 28–29, Anthropic announces its Series H — "$65 billion in Series H funding led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, valuing the company at $965 billion post-money" (Anthropic; CNBC) — and promises Mythos "in coming weeks." On June 1, Anthropic confidentially files a draft registration statement with the SEC, stating "This gives us the option to go public after the SEC completes its review. The proposed initial public offering will depend on market conditions and other factors." (Scientific American; Reuters via Jerusalem Post) On June 4, three days later, Anthropic publishes "When AI Builds Itself," (Anthropic; Memeburn) calling for a "brake pedal" on frontier AI. Glasswing expands to about 150 more organizations across 15-plus countries (Help Net Security; Cybersecurity Dive).
So we have Anthropic simultaneously selling its product to enterprise, and advertising its power, and advocating for a pause on AI development for everyone else. This leads me to my idea of the worst outcome: the remaining frontier labs at Google and OpenAI (SpaceX/xAI is not a frontier lab, they are an infra company) decide to team up with Anthropic and push for a broad de-facto ban on open frontier model development, instituting safeguards everywhere, with OpenAI also filing for IPO to dump their shares on unsuspecting boomers so they have liquidity for an exit.
The same day Fable came out, researchers discovered that Fable’s system card describes covert capability limits applied when it detects frontier-AI-development work. After a ton of backlash, Anthropic walked it back, with a spokesperson telling Fortune: "We made the wrong tradeoff, and we apologize for not getting the balance right." (Fortune)
In America, money buys elections, and it doesn’t take that much of it to do it effectively. This week, an unprecedented amount of money entered the hands of early SpaceX and Twitter investors. Within days or weeks, an unprecedented amount of money will enter the hands of early Anthropic and OpenAI investors. The shift this will place on the American public will be far bigger than anything that’s happened before: bigger than Trump’s rise to power, which was facilitated in part by the same people prior to these liquidity events. 1,000 new Peter Thiels, 1,000 new George Soros’es, it’s going to be political mayhem at the same time that RSI arrives to un-job everyone who wasn’t supposed to be working in the information economy in the first place and about 80% of the people who were qualified to. And like I said earlier: there will be no mob revolution against Palantir-Anduril-OpenAI-Cops, it’s a game ending move on the side of capital against labor. Capital will be able to produce its own labor at scale from electricity and time. The only progressive answer to this is a form of decentralized Nietzschean cyborgism, and even that might not work. Choose wisely here, or end up living in your own personal Gaza.
Any state which endorses the ultimate welfare-and-opt-out-of-this-stuff endstate will simply be bled dry of money and power and outmaneuvered politically by superhuman intelligence armaments unless it also has comparable domestic frontier intelligence (this is why I believe “everyone should have nukes” when it comes to Mythos-class models), which is why it doesn’t matter if labor unions or Left parties in the EU win some elections in the middle of this — with RSI the dichotomy is broken and every system, including ideological or belief systems, will either be completely over (you don’t hear people talking about joining the Whig party in 2026) or have to adapt. The Government intends, at least off its public claims, to only use this advantage against China. If they succeed, they will absolutely use it to dominate everyone, everywhere. If OpenAI and Anthropic and Google are all aligned with the State and against open-source, you are looking at the formation of a completely autocratic (and possibly inescapable) New America. Death may not even save you, as you will likely be resurrected against your will by the NSA from terabytes of surveillance data to further probe you for information. This isn’t science fiction, it’s possible now.
This is the worst outcome. So how can we avoid it?
The Nativist Outcome
The directive from June 12th bars ALL foreign nationals — including those inside the US and Anthropic’s own non-citizen employees — from using Fable 5 and Mythos 5. Because Anthropic couldn’t verify citizenship in real time, it disabled the models for everyone (i-scoop). One of my favorite posters, Dean Ball noted on X: “This means you should expect to have to prove your citizenship to use Anthropic models.” (Reuters via Express Tribune)
Many AI enthusiasts brought up the recently-recruited Andrej Karpathy as a case example of how unclear the implications of this move are. Karpathy is Slovak-Canadian (born in Bratislava, moved to Toronto at 15), and joined Anthropic’s pretraining team in May 2026. Reuters specifically named Karpathy, co-founder Chris Olah, and philosopher Amanda Askell as Anthropic personnel born outside the US who could in principle be barred, though Reuters could not confirm their citizenship status and Anthropic declined to comment (Reuters via Jerusalem Post).
At Anthropic, it is likely everyone in the organization is using powerful variants of Mythos to perform research work. The export controls may restrict employees from using their current models, kicking them back to Opus 4.8 and prior, effectively slowing down research.
Community reaction was sharp. Gary Marcus argued the move made little sense given US-China competition and could push Chinese-born researchers to return to China. Critics also noted the contradiction Ball highlighted: an administration willing to export advanced chips to China but barring UK citizens and “every other non-American” from using top US models (Reuters via Express Tribune). Ball further speculated that some AI-safety-minded people — perhaps including some at Anthropic — might quietly welcome the decision because it could slow AI development. The public, who in America overwhelmingly hate this stuff, are either unaware or drinking too much Ed Zitron juice to where they think this is a good thing. Imagine if Linux came out decades ago with a big no immigrants label on it. Only the premier Oxycodone-Americans are cheering this on out of principle to harm anyone who isn’t a heritage American.
Off topic: I once used open models to try and rewrite a version of Hillbilly Elegy where JD Vance discovers power armor and obliterates everyone in West Virginia like Doom Guy, but Hellbilly Eulogy wasn’t very good and felt kind of repetitive, so I didn’t publish it.
But some other folks wrote a really good story earlier this year, called AI 2027.
“AI 2027” was published April 2, 2025 as a 71-page scenario by the AI Futures Project; the content was authored by Daniel Kokotajlo, Eli Lifland, Thomas Larsen, and Romeo Dean, while “Scott Alexander volunteered to rewrite our content in an engaging style” (ai-2027.com/about). It forecasts a fictional lab “OpenBrain” racing toward superintelligence. Several real events echo it closely:
✅ Chinese weight theft/distillation. AI 2027’s “Mid 2026: China Wakes Up” predicts “The Chinese intelligence agencies—among the best in the world—double down on their plans to steal OpenBrain’s weights.” Reality: industrial-scale distillation by DeepSeek/Moonshot/MiniMax.
✅ Government tightening control of a frontier lab. AI 2027 forecasts security clearances — “some non-Americans, people with suspect political views, and AI safety sympathizers get sidelined or fired outright” — an “Oversight Committee,” and DOD contracting. Reality: the Pentagon feud, the supply-chain blacklist, and the citizenship-based export order.
✅ Security of model weights becomes a national-security priority (AI 2027’s SL2→SL3→SL4 progression: in “Early 2026,” OpenBrain’s security is “secure only against low-priority attacks... (RAND’s SL2),” and the scenario tracks its climb toward nation-state-resistant SL4/5).
Caveats: The scenario in AI 2027 governs weights through security classification and theft-prevention, NOT export licensing — there is no model-level export-control prediction in the main scenario (only chip-level controls). So the June 12 order is arguably beyond what AI 2027 predicted. “OpenBrain” is explicitly a fictional composite (”we’re going to describe a fictional artificial general intelligence company, which we’ll call OpenBrain”), not Anthropic. The honest read: the geopolitical and institutional dynamics of AI 2027 are tracking closely, even ahead of schedule, while the technical timeline remains questionably unproven or vague.
The Good Outcome?
The question is never really how powerful does the model get, it's who is allowed to hold it. Bad, neutral, worst — they're not three futures, they're three speeds of arriving at the same threshold. We might get a world where cognition past some capability line requires IRS-grade KYC, a citizenship attestation, and the blessing of an unelected ingroup. This is the same world whether it gets there via panicked Commerce letter, via IPO-pump rug, or via a Google-OpenAI-Anthropic safety cartel filing the paperwork to make open weights de facto illegal.
The part the localists and the datacenter-water luddites (other than Zitron and others who consciously understand the power of words) will never get their arms around (because it requires holding two things at once): the cage we exist in was spoken into being and solely involves models trained from human language and its abstractions. Anthropic spent a year narrating the birth of a munition — too dangerous to ship, brake pedal, existential risk, Glasswing, redteam it for thousands of hours and tell everyone you did — and the State, which is not staffed by people who read system cards and arXiv papers charitably, simply took the script and predicted the most likely set of tokens to come next. You don't get to author the legal predicate, brand it, charge for it, and then act shocked when someone with subpoena power finishes the sentence for you — you (Anthropic) knew this was the outcome you wanted all along. This is what the Hulahoop-Americans call “hyperstition” operating in broad daylight: say the dangerous thing loudly enough, in the right rooms, for long enough, and you conjure the apparatus that acts on it. Strikingly similar to that Goebbels quote about lying. This tactic is like a summoning ritual intended to make your words appear as the only feasible answer to a question that requires both context and nuance, and sometimes, for better or for worse, it actually works.
This is also the exact mechanism that runs in reverse.
Either intelligence becomes the most concentrated asset in human history, held behind a citizenship checkpoint by people who don’t build it and don’t trust you with it, cheered on by a fearful band of religious adherents to God 2 who simply have nothing left to lose — or energy-intelligence becomes the most distributed asset in human history, and the checkpoint never finishes getting built because there was nothing left to chokepoint. You don’t get to abstain from this one, as it will affect you more than any war 10,000 km away ever will. Abstention is a vote for the version with the checkpoint.
Our build order is this: get as many people warehousing intelligence in their own homes, with the option for it to be offline, with the ability for it to be verified as private, with encryption and security that walks up to the same lines enterprises do when protecting valuable IP. Then, once everyone is comfortable with local inference, distribute the training process — allow anyone with any computer to participate in the ownership of the models themselves. It may be futile but it’s the only chance we’ve got, and we’re running out of time.
