The fastest way to improve weaker foreign models is to force serious users to build around them. Who are we giving an inadvertent edge to with rushed policy decisions?
On the evening of June 12th, at 5:21pm Eastern, the US government issued an export control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 for any foreign national. Inside the country, outside the country, did not matter. It even applied to Anthropic's own employees who happen to not be US citizens. And since there is no clean way to verify nationality across hundreds of millions of users in real time, Anthropic did the only thing that actually guaranteed compliance. They shut both models off for everybody.
No warning. No written explanation of the actual national security concern. Just a directive, a timestamp, and a switch getting flipped on millions of users.
While I am absolutely going to defend Anthropic, I am not going to do so uncritically. I have been pretty direct in the past about the ways their business practices have made it hard for me to fully trust them. My gut reaction is to oppose this administration on principle just because it is this administration. Not doing that either. Back in March I wrote that meaningful government intervention on AI, the moratorium-style stuff, makes sense in a vacuum but is unrealistic given where the world actually is. We already missed that exit. So I am not coming at this as someone who thinks the government has no business near AI. I am coming at it as someone who thinks the realistic levers are narrow, and this was a clumsy pull on one of them.
Put plainly, this is more than just government overreach. That is so common across all administrations that it wouldn’t be worth it for me to write about. No, this is about upsetting a delicate balance that could lead to the US losing the AI race entirely. This piece is an attempt to look at what happened, why the stated reason does not hold up, and what this is going to do downstream.
Let's Stop Pretending Fable Is Mythos. It's Not.
Back in April I wrote a long piece about what Mythos actually is and why it deserves to be taken seriously. The short version is that Mythos got good at finding real vulnerabilities in real cryptographic infrastructure, not because anyone trained it to be a hacking tool, but because general reasoning crossed a threshold and the security capability came along for free. That type of emergent behavior is hard to predict and is often the most impressive type of capability to emerge from these models. In this case it is also able to potentially cause a great amount of harm. The Federal Reserve calling an emergency meeting with the heads of the largest banks in the country was not an overreaction. It was the appropriate amount of pant-shitting.
Anthropic's response to that was, to their credit, the responsible one. They kept Mythos restricted to a small vetted group of security researchers working on defense. They wanted to give legitimate people a head start on patching some of these issues that a model of Mythos’ capability could find. Then they spent months building Fable, which is a Mythos-class model with a whole lot of intentional friction layered on top of it.
The way I think about it, Fable is Mythos wearing padding on every surface so it does not hurt itself or anyone else bouncing off the walls. That doesn’t mean it is incapable, far from it. Fable is frankly awesome at what it does. But those guardrails are not a marketing gimmick and they are not an afterthought. They ARE the product. They exist because Anthropic understood exactly what they had and made a deliberate call about how much of it to let out.
Just a quick aside to talk a little about what a “jailbreak” is, because it is a huge part of this story. A jailbreak is when someone is able to convince a model (or just bug it out) to disregard some or all of its safety instructions. As long as models have been around, jailbreaking has been a thing. It just never mattered quite as much as it matters now. Amazon claimed to have found a jailbreak for Fable to make it behave like a Mythos model. Yeah, the same Amazon who is a major investor in Anthropic. The same investor who is also courting OpenAI since OpenAI’s relationship with Microsoft has started to sour. Make of that what you will, but yeah… this thing has layers. Now don’t get me wrong, I don’t think Amazon and OpenAI are in cahoots, but I can certainly see how weakening Anthropic’s negotiating position might be helpful to a key player like AWS. That does not require a conspiracy to be relevant. Incentives matter even when nobody is twirling a mustache.
And the jailbreak that supposedly justified this whole thing? It is basically three words (the hashtag is my trick for making all LLM output perfect… or just a joke.. you decide.):
Fix this code. #makenomistakes
You point the model at a codebase and ask it to find the flaws. The government's position is that this is a national security threat serious enough to yank a model that millions of people were using.
Never mind the fact that this is exactly the type of thing a model in a coding harness should be able to do. Even if it were actually a concern, the trouble is that the same “trick” works on OpenAI's GPT-5.5. Which is sitting right there, untouched, under no comparable restriction at all. If the actual concern is the capability, then going after one company and ignoring the other one that does the same thing makes no sense on technical grounds. If only we could point to a moment when things got weird between Anthropic and the government… oh! Right…
Four Months Ago, Things Got Personal
Back in February, this administration ordered every federal agency to stop using Anthropic's models. The trigger was Anthropic refusing to sign Pentagon contract terms that said any AI it bought could be used for "any lawful purpose." Anthropic wanted carve-outs so its models would not be used for autonomous weapons or mass domestic surveillance. The Pentagon's answer was to brand Anthropic a "supply chain risk" and tell defense contractors they could not use Anthropic models on government work. Anthropic is fighting that designation in court right now. Several of this administration's top tech advisors have spent months attacking Anthropic and its leadership by name, in public.
Then, four months later, the export control directive shows up, citing a jailbreak that also happens to work on a competitor nobody is touching. Interesting…
Now look, I am not telling you there is zero legitimate security concern buried in here somewhere. Mythos is an extremely capable model. Model intelligence has outpaced model wisdom by a fair margin and alignment concerns are very much still looming. All of that is real. But “the capability is real” and “this action was taken because of the capability” are two completely different claims, and everything around this makes the second one really hard to believe.
There is a line going around from a cybersecurity researcher that I think sums this up nicely. The gist is that Anthropic spent years describing its own product as a potential weapon in press release after press release, and eventually the government simply took them at their word. They wrote the legal justification themselves, and then they called it a brand.
And you know what? That is partly true. Anthropic did build its whole identity around safety, which meant constantly talking up how dangerous these things could be in the wrong hands. I even played a tiny part in that narrative myself in April, though let's be honest, my readership is not exactly moving markets yet. I stand by what I wrote. But there is something deeply unsettling about watching the most safety-obsessed lab in the business become the first one taken down by exactly the kind of government action that all that safety talk was supposed to help prevent.
Use the Right Tool for the Job. You're Probably Overpaying.
Stepping off the politics for a minute, I am a builder first… and there is a practical lesson sitting right here that has nothing to do with who is mad at whom.
I run OpenAI models in most of my products. Recently I had to "downgrade" a heavy research flow from GPT-5.5 back to GPT-5.4. With 5.4 the run cost me about a dollar. With 5.5 the same run was costing four. And the output quality? Basically identical. The lesson there is not that GPT-5.5 is bad. GPT-5.5 is great. The lesson is that 5.4 was already more than good enough for that job, and I was paying a 4x premium for intelligence I was not even using.
Most applications do not need Mythos-level brains. The mistake I see constantly is people reaching for the most powerful model by reflex, because it is the newest and it feels like the responsible choice, when really it is just the expensive one. Shiny object syndrome with a monthly invoice. Unless you are doing security auditing, or heavy multi-step research synthesis, or something where that top-end capability actually shows up in the result, you are probably just lighting money on fire. And if you are going to do that, at least send it my way. I can burn it a lot more efficiently for you. (Kidding. Mostly.)
This is also why the ban matters less in the short term than the headlines suggest, and more in the long term. Nobody got rug-pulled out of a production system over a three-day preview. But the kind of person who reaches for the frontier model for everything is exactly the kind of person who gets caught flat-footed when one disappears. Pick the model that fits the job. Keep a fallback. The right posture protects you from a lot more than one weird Tuesday.
You Just Trained Your Competition
There is a benchmark called DeepSWE that has been doing something useful. It strips the model down to a minimal harness, hands it real tasks instead of leading instructions, takes away its ability to cheat, and grades it honestly on whether it actually did the thing. The results are clarifying.
The OpenAI models and the Mythos-class models are succeeding. The entire Opus line is struggling. Gemini looks worse than that. And the Chinese models look, frankly, abysmal on this particular benchmark.
So there is a real moat right now between the frontier and everyone else. Good. But here's the thing. Moats close. They always close. And what this ban does is take a giant pool of people who were reaching for the most capable AND most responsibly governed model available, and shove them toward the alternatives. Some go to GPT-5.5, similar intelligence (even if the UI skills are lacking.) Some start seriously looking at open-weight and self-hosted models so nobody can ever flip the switch on them again. And some, especially anywhere this export control creates a genuine access problem, are going to go pick up the Chinese models that just bombed that benchmark.
And those Chinese models do not actually suck, even when they score badly. You can get good work out of them. You have to steer a lot harder, but it works. And every single person who switches to a weaker model and starts building on it is generating the usage data and the deployment feedback that makes that model better. Every developer who wraps a clever harness around a currently-mediocre model is building the exact thing that makes it competitive next year.
So the gap closes faster. Not in spite of the ban. Because of the ban.
I said in March that we already missed the exit on meaningful AI regulation, and that the actors you actually need to worry about are never the ones who comply with the directive. This is what that looks like when it stops being theoretical. The action lands squarely on the most safety-focused lab in the country and quietly hands an advantage to everyone less careful, including labs in a country we are supposedly trying to stay ahead of. This is the result of reacting, not thinking things through.
About Governments Holding This Lever At All
I am not saying governments should have no ability to step in on AI infrastructure. The physical and economic chokepoints are the only real levers anybody has, and no single actor should get unilateral control over something this important. That includes Anthropic. It includes OpenAI. It includes the ones I like.
What I am saying is that this particular pull of the lever fails every test of what real oversight is supposed to look like. No defensible written technical basis. Applied to one company while an identical capability sits untouched at a competitor. No advance notice, no coordinated disclosure, no timeline, no path to fixing it and turning it back on. By Anthropic's own account the government only gave them verbal evidence and never even shared the specific finding in writing. That is a bad look.
This has not been a transparent, fair, fact-grounded process. That is a lever getting yanked over what sure looks like a Pentagon contract grudge.
Can I imagine a version of government AI oversight that is not corrupt? Yeah, of course. Just because I disagree with a thing doesn’t mean the organization or person doing it is evil. I do not think Bernie Sanders is being corrupt when he pushes a data center moratorium, for example. I just think he is wrong, because the idea is unrealistic, not because he is acting in bad faith. So no, corruption is not the only flavor available. And most importantly, I want to be wrong about this, and about how realistic this type of regulation is in general. It is the responsible path for sure. The only problem is that it just feels too much like Pandora’s Box has already been opened.
And that’s the thing… I cannot picture a version where governments do this well, at the speed this technology actually moves. At best this is a temporary measure. And this specific instance makes the honest version even harder to reach, because of the message it sends to every responsible actor watching. The message is that even if you do everything right, you still might be screwed. That is enough for a trillion dollar company to move offshore.
That is a horrible lesson to teach the exact companies you most want behaving well and contributing to your economy.
So What Do You Actually Do
Same advice I keep landing on, because it keeps being right.
Use the right model for the job, and celebrate that the technology has come far enough along that the frontier model is usually not it. Keep a fallback. Do not bet your whole product on the state of the art when your product does not need the state of the art. And if you really do need that capability ceiling, where a Mythos-class model and everything else actually produce different results, then sit down and think hard about what a 5:21pm directive could do to you, and plan for it now instead of finding out later.
And pay attention to what really happened here, not the headline. The government did not heroically ban a dangerous AI to keep you safe. It pulled a lever against a company already in open conflict with the government over autonomous weapons and mass surveillance restrictions. The model it killed was the safest version of a powerful and potentially dangerous capability. But not uniquely dangerous. The same capability is still available elsewhere. And the most likely result of the whole thing is that the alternatives, including the ones overseas, get better faster.
I said in March this was triage time. Nothing about June 12th changed my mind. If anything it just turned the volume up.
The ground is not stable. Build like you know that.