Own or Be Owned: The Week a Letter Switched Off Fable 5

I coded for seven hours on Fable 5 the day before it disappeared.

It felt like the best tool I had ever used.

The next night it was unreachable.

Not a bug. Not a price change.

A decision made by people I will never meet - that the company who built it could not override.

Here is what actually happened, because most of the takes skipped it.

A letter at 5:21pm

On June 12, at 5:21pm Eastern, the Trump administration sent Anthropic an export-control directive citing national security.

It ordered the company to cut off Fable 5 and Mythos 5 - their two most advanced models - from any foreign national, inside or outside the United States.

That included Anthropic's own foreign-born employees.

Anthropic couldn't separate foreign users from American ones fast enough to comply.

So they did the only thing they could: they disabled both models for everyone, worldwide.

A letter arrived at 5:21pm. By the evening, the model was gone.

Whether it was justified is not the lesson

The government says it acted on a jailbreak.

The security expert who reviewed the actual evidence called it "a complete overreaction" - the ordinary kind of red-team prompting defenders use to patch their own systems.

Anthropic and this administration have been fighting for months over AI guardrails.

So this may be less about safety than about a fight neither side will explain in public.

I have a view on who was right. It doesn't matter. That's the whole point.

It didn't matter what I thought.

It didn't matter what Anthropic thought - and they built the thing.

The model that felt indispensable on Thursday was unreachable on Friday.

Nobody in the transaction I had with it - not me, not the company I was paying - had the authority to keep it on.

If you don't own the model, you don't control your access to it.

You're renting. And the landlord has a landlord.

This week, the landlord's landlord sent a letter.

This wasn't even the first warning that week

Days earlier, Anthropic had shipped Fable 5 with a quiet defect researchers called "secret sabotage": the model silently gave worse answers if it detected you were working on frontier AI.

No notice, no flag - just degraded output.

Two days of backlash later they apologized and made the behavior visible. Credit for the fast reversal.

But the story was never the apology.

The story is that a vendor can change what your AI does, silently, and you might never know.

And most of us can't simply leave when they do.

Switching models is easy. Switching everything around the model is not.

Your prompts are tuned to one model's quirks. Your embeddings don't transfer. Your agent's memory and context live on their servers.

A mid-size team burns two to four weeks of engineering and $15-30K to migrate - and ships zero new features the whole time.

That's the real lock-in. Not the model - the plumbing.

The careful ones did their best. It lasted 72 hours.

Here is what should unsettle even the optimists.

Anthropic is, genuinely, one of the most safety-obsessed labs on earth.

Its CEO talks about wanting checks and balances everywhere, about treating Oppenheimer as a warning rather than a model to follow.

Early testers reportedly told him a version of "this is a super weapon, please don't release it."

Now look at one week.

The most careful lab in the world shipped a model testers called a super weapon.

Researchers jailbroke it within a day.

A government letter took it off the board inside 72 hours.

If the most careful company on the planet cannot keep its flagship online for three days, the safety story the whole industry tells you is not a plan.

It's a hope.

The technology ships faster than anyone can govern it - including the people building it.

That is not a reason to stop using AI.

It's a reason to stop assuming someone else has it under control.

The number that reframes everything

There's a figure I keep coming back to.

Anthropic pays roughly $1.25 billion a month to rent compute from a single vendor, drawing more than 300 megawatts from one facility today.

Canada's entire sovereign-compute and national-AI commitments add up to something like $5 billion over five years.

About three months of one American company's rent.

I'm not knocking the strategy. It's real progress.

The point is subtler, and it's the same point the Fable shutdown made in a different key.

At the frontier, access isn't a pricing problem. You cannot simply pay more and get more.

Access is a capacity-allocation decision made by someone else, in another country, under their own constraints.

That's a fine arrangement right up until their constraints stop aligning with yours.

The entire question of sovereignty is what happens when they don't.

The uncomfortable irony

While the American models were going dark, the most reliable models in the world were suddenly the open-weight ones - many of them Chinese.

DeepSeek, Qwen, MiniMax.

You download them. You run them.

No letter reaches into your own machine and switches them off.

I'm not celebrating that.

It's a strange world where the models you can most depend on are the ones no government can revoke.

But it's the world we're in now - and pretending otherwise is how you get caught with your production system dark.

Rent the intelligence. Own the system.

So here is the position I've landed on, and it isn't ideological - it's operational.

I pay for Claude. I'll keep paying. It's the best model I've used.

But my sensitive workflows run on hardware I control.

And every workflow I build, I build so the model is swappable.

Rent the intelligence. Own the system around it.

That is the only stance that survives your vendor's next bad tradeoff - or your vendor's government's next letter.

This is the entire reason we are building SimpleDirect®: open models for Canadian work, downloadable, yours.

Not because owning is cheaper - it usually isn't.

Because rented intelligence can be revoked by someone you have never met, for reasons they may never explain.

I said a version of this for a year and people thought it was a bit much.

Own or be owned.

That stopped being a slogan at 5:21pm on Friday.