The landlord's landlord pulled the plug
I coded for seven hours with Fable 5 on a new personal project last Thursday.
It felt productive.
It felt like the future.
Then Anthropic got a letter from the US government at 5:21 p.m.
ET on Friday, and by the weekend Fable 5 and Mythos 5 were gone for everyone, everywhere.
Not throttled.
Not geofenced.
Off.
Goodnight, Fable 5.
It was nice knowing you.
That sounds sentimental, and it is, but the sentiment is not the point.
The point is what this episode just taught every founder, developer, and government on earth about who actually controls frontier AI.
What happened
Anthropic said it received a US export-control directive that barred access to Fable 5 and Mythos 5 for foreign nationals, including foreign national employees inside the United States.
Read that again.
Foreign nationals inside the US.
People sitting in San Francisco offices.
That is the part that makes this unprecedented.
You cannot comply with that by blocking IP addresses or geofencing a region.
There is no clean technical way to exclude only certain people from a model API, so Anthropic did the only thing it could and shut both models off globally.
The Commerce Department letter reportedly required licenses for the export, re-export, or even domestic transfer of the models.
Other Claude models were untouched.
Anthropic also said the government never spelled out the specific national-security concern, though the company believed officials were worried about a potential jailbreak technique against Fable 5, one it described as narrow and based on already-known vulnerabilities.
So the company's newest, most capable frontier models were pulled back almost the moment they shipped.
Why this is different from every AI policy story before it
We have spent two years arguing about AI policy as if it were a debate.
Safety frameworks.
Voluntary commitments.
Compute thresholds.
Red-teaming reports.
This is not a debate.
This is enforcement, and it reached inside the building.
If the concern is a jailbreak rather than stolen weights, the government just treated operational misuse of a model as a national-security event.
That is a categorically different thing than worrying about someone leaking model weights to a foreign adversary.
It means a model's behavior can now be the trigger for a nationality-based access restriction.
And if that becomes a template, global rollout of any frontier model stops being a product decision and becomes an export-review outcome.
Product readiness no longer determines whether the world gets access.
A letter on a Friday afternoon does.
I keep coming back to the cruelty of the sequencing.
Had Fable 5 been gated from day one, we would have shrugged and moved on.
The worst part is giving you a taste of the lollipop and then taking it away.
They let us build with it.
They let us feel productive.
Then they pulled it.
The lesson nobody wants to say out loud
Here is the uncomfortable truth this week forced into the open.
If you rely on frontier proprietary models, you are not a customer.
You are a tenant.
Fable 5 was leaving my own subscription on June 22 anyway.
I was literally about to buy API tokens to keep building.
Now I will spend that money somewhere else, because why would I pour a project's foundation onto ground I do not control?
I coded with Opus afterward to keep moving.
It is still a good frontier model.
But honestly it did not feel as good as Fable, and that almost doesn't matter anymore.
The model itself was never the real question.
Access is the real question.
The landlord's landlord just pulled the plug, and there was nothing Anthropic, the actual provider, could do about it.
They are tenants too.
Own or be owned.
I have said versions of this for a while, and I have never felt it as concretely as I did this weekend watching two models I was actively using vanish overnight.
What my friends and I are actually doing about it
This is not a thought experiment for me.
My friends and I are now seriously considering training our own frontier AI on top of strong open models like DeepSeek and MiniMax.
Is that expensive?
Yes.
Insanely so.
But here is the thing people miss: this was our plan before the plug got pulled.
The shutdown just removed the last excuse to keep renting.
And with SimpleDirect® growing, we have more compute and more leverage to actually pursue it than someone whose entire AI stack is rented frontier access ever could.
That is the asymmetry I want every founder to internalize.
Ownership compounds.
Renting does not.
When you rent, every dollar you spend buys you exactly one month of access and zero equity in your own future.
When you own the weights, the fine-tunes, the infrastructure, every dollar builds something that survives the next letter on the next Friday.
Open weights are not just cheaper in some scenarios.
They are yours.
They can be locally hosted.
They sit outside the central access controls that just nuked Fable 5 in a single afternoon.
I hope we can make this right.
But I am not waiting around to find out whether we can.
The pattern: everyone is for themselves now
Here is the part that is bigger than Anthropic, bigger than the US, bigger than this one week.
China and their DeepSeeks, MiniMaxs, and Qwens are laughing right now.
While US export pressure tightens around closed frontier models, the Chinese open-weight strategy keeps spreading models that anyone can download and run.
You cannot send an export-control letter to a model that is already on ten thousand hard drives.
That collision, US controls on closed models versus China's open diffusion, is the real story under the headline.
But watch what it does next.
After this, every country starts its own AI-building war.
It had already started, but it was a two-player game between Washington and Beijing.
Now it spreads.
The West has to seriously rethink relying on the US for frontier AI, because if access can be revoked for foreign nationals inside America, no allied country can build critical infrastructure on top of it with a straight face.
And China's allies may start asking the same question in reverse.
Do you want your national AI stack dependent on Beijing's release decisions?
The middle countries are the ones caught hardest.
India, the Gulf states, and others now face a forced choice: pick a side, or build their own.
The rational answer, for almost all of them, is build your own.
Which means compute gets harder to grab worldwide.
More money gets burned on duplicated effort.
And AI access fragments further than anyone selling the "one global model" dream wanted to admit.
Everyone is for themselves now.
What I'm watching
First, whether Anthropic gets any waiver or narrowed scope from Commerce that lets Fable 5 and Mythos 5 come back even partially.
We may genuinely have seen the last of those two models.
Second, whether the government ever makes the underlying concern explicit.
If it really is a known, narrow jailbreak, the precedent is doing far more damage than the vulnerability ever could.
Third, whether other frontier labs get the same treatment.
If OpenAI or Google ships something judged comparably dual-use, does the same letter land on their desk at 5:21 on a Friday?
And fourth, whether developers displaced by this keep flowing toward open weights they can host themselves.
Every one of those moves is a small permanent loss for the closed-model business model.
Close
I will miss Fable 5.
It was a genuinely good tool and I did good work with it.
But the kindest thing it ever did was disappear, because it showed me exactly how little I owned of the thing I was building on.
The next great AI companies will not be the ones with the best access.
They will be the ones who never needed permission.
Sources: businessinsider.com, fortune.com, techtimes.com
| 
