I found my groove with Claude Fable 5. Then the government took it.
No warning. A model I had built my mornings around, pulled overnight by federal order, over a flaw its biggest competitor ships freely and faces no penalty for. I want to tell you exactly what I lost, because I measured it before they took it, and the number is the part that should worry you.
I build in the mornings. In a terminal, inside the operating system I run my businesses on. So when Anthropic released Fable 5 on June 9, I didn't read about it. I put it to work.
Then I did what I tell every client to do before they trust a tool with anything that matters. I measured it.
I Did Not Trust the Hype. I Ran the Numbers.
Over four days I ran Fable against the model I'd been using, Claude Opus, across the actual work I do: building features and fixing bugs in my own codebase. Then I split the sessions by the kind of work, so I was comparing building to building and fixing to fixing, not apples to oranges. Here is what my own data said.
On feature builds, twelve Fable sessions against forty-four on Opus: Fable averaged 58 minutes of active work per session, Opus 69. Same output per instruction. Same reply depth. I corrected Fable on one percent of turns, Opus on three. We hit the same number of test failures along the way. Ninety-two percent of my Fable sessions ended in shipped, committed code. Eighty-nine percent of the Opus ones did.
On bug fixes, seven Fable sessions against sixteen: Fable averaged 22 minutes, Opus 33.
Read past the decimals and the finding is simple. Same quality. Roughly 15 to 30 percent less of my time, depending on the work.
I'll name the soft spot in my own analysis, because the honesty is the point. All of my Fable use landed late in the project, when I knew the codebase cold and had momentum. Some of that speed edge is the model. Some of it might be me. To rule that out completely I'd have needed to run both models on the same days on the same tasks, and I mostly didn't. So I hold the number loosely.
But I held it long enough to make a decision. For the building and fixing I actually do, Fable gave me the same result faster. So I switched my default. That is what a tool is for.
Then it was gone.
The Letter Arrived at 5:21 on a Friday
On June 12, at 5:21pm Eastern, Anthropic received an export control directive from the US government, sent through Commerce Secretary Howard Lutnick. It ordered the company to suspend all access to Fable 5 and its more powerful sibling, Mythos 5, for any foreign national, anywhere in the world, including Anthropic's own foreign-born employees.
Because there was no clean way to block only some users, Anthropic shut both models off for everyone. Every customer. Worldwide. Its other models stayed live, which is why I could fall back to Opus at all. This appears to be the first time a leading AI company has pulled a publicly deployed model off the market because the federal government told it to.
Here is the reason the government gave, as far as anyone can tell. The letter cited national security and offered no specifics. Anthropic's understanding is that someone demonstrated a way to bypass Fable's safeguards. The technique, by Anthropic's account, amounts to asking the model to read a body of code and identify its flaws. The company reviewed it and found the vulnerabilities minor, simple, and already discoverable by other models available to the public right now.
Including, by Anthropic's own statement, OpenAI's GPT-5.5.
Which faces no comparable export controls.
Sit with the shape of that. The capability the government called a national security threat in one company's model is, according to the company that built it, sitting inside a competitor's model that remains fully available to anyone who wants it.
It gets harder to read as a neutral safety decision the longer you look. Reporting indicates the administration had already tried to stop Anthropic from releasing this model and failed, then moved after a different company claimed it could jailbreak Mythos. Anthropic, for its part, has publicly disagreed with the order, called it a likely misunderstanding, and said it is preparing to release evidence. The company's position is direct: recalling a model used by hundreds of millions over one narrow flaw, applied as a standard across the industry, would halt new model releases for everyone.
I am not going to tell you what was in anyone's head. I am going to tell you what is on the record and let you do the math. A government moved against one company's tool, citing a risk that its competitor's tool carries freely, after failing to block that tool through other means, while the company being targeted says the stated reason does not hold up.
You can decide for yourself what that is. I have decided what it is.
You Priced Vendor Risk. You Did Not Price This.
Here is my position, and it is the reason I am writing instead of stewing.
Every operator who depends on a vendor has learned to price three risks. The price goes up. The product gets discontinued. The service goes down. You have lived all three. You have a clause in a contract or a line in a budget for each of them.
What happened on June 12 is a fourth risk, and almost nobody had it on the board. Your vendor's government.
A tool you depend on can be taken from you overnight by a third party who is not your vendor, who holds no contract with you, who is not accountable to you, for reasons that may have nothing to do with you and may not even be true, while your vendor stands beside you saying the decision is wrong and is powerless to reverse it.
You cannot negotiate with that risk. You cannot write your way out of it in a service agreement. It sits above the entire relationship you thought you understood.
And it is not going back in the box. The moment a government shows it will reach past a company and pull a deployed model, every leader building on AI is operating in a different environment than they were a week ago. Microsoft had already moved to restrict its own employees' access to Fable before the order even landed. The largest technology companies in the world are treating model access as something that can be revoked. You should too.
Owning Your Architecture Made Me Survivable. It Did Not Make Me Immune.
I have built my company on a principle I repeat to every client. Own your architecture. Rent the engine. Your processes, your documented workflows, your content, your systems live in infrastructure you control, not inside any one AI platform. The model is the engine. The engine is supposed to be swappable.
On Saturday morning, that principle paid off in the most literal way possible. I changed one line, the name of the model, from Fable to Opus. My skills, my workflows, my content, my systems all kept running. I lost no work. The version of me who had built everything natively inside one model's quirks spent the weekend rebuilding. I spent it shipping.
So the principle held. And here is where I have to be honest with you, including correcting something in how I have taught this.
"Rent the engine" carries a quiet assumption: that engines are interchangeable. That one is as good as the next. That the swap costs you nothing.
That assumption is wrong, and my own data is the proof. The engine I swapped to is a real step down on the one axis I had measured and chosen for. Not in quality. In my time. Fifteen to thirty percent of it, every session, indefinitely, with no say in the matter and no timeline for getting the better option back.
Owning my architecture did not make me whole. It made me survivable. Those are different things, and the gap between them is the entire lesson.
There is one more honest corner I am standing in, and I would rather name it than have you catch it. The single hedge that would have fully protected me is a fallback to a model from a different company. I will not take it. I chose Anthropic deliberately, on ethics, after watching how different companies in this space behave, and I am staying. Which means I am knowingly exposed. I traded a measure of resilience for a value I am not willing to sell. That is a real cost, and pretending otherwise would make me exactly the kind of consultant I do not trust.
What a Serious Leader Audits This Week
If you run an organization that depends on AI for anything that matters, here is the work in front of you. None of it is panic. All of it is overdue.
Find out where you are single-threaded. Every place your operation leans on one specific model from one specific company with no tested fallback is a place a letter on a Friday can stop you. You do not have to eliminate that. You have to know where it is and decide on purpose, the way I did, rather than discover it in an outage.
Make sure what is yours is actually yours. Your workflows, your prompts, your documented processes, your data. Are they sitting in infrastructure you control, or inside a platform that could change its terms, or disappear, tomorrow? If you cannot redeploy everything you have built onto a different engine in an afternoon, you do not own your architecture. You are renting all of it.
Build the fallback you can live with. For me, that is another Anthropic model, because my line is the company, not the specific model. For you it might be a second provider entirely. The choice is not the point. The point is that you have made one, tested it, and know exactly what it costs you in quality or speed before you are forced to find out live.
And price the fourth risk. Put it on the board next to the other three. Not because you can prevent it. Because leaders who can name a risk make better decisions than leaders who get surprised by one.
The Tool Comes Back or It Doesn't. The Lesson Stands Either Way.
By the time you read this, Fable may be back. Anthropic is fighting the order and may win. I hope they do, partly because I want my fifteen percent back, and mostly because the precedent is worse than the inconvenience.
But whether or not I get the tool back, the week taught me something I will not unlearn. I did everything right. I chose my tools on principle. I measured before I trusted. I built my own architecture so I would never be trapped inside someone else's. And I still woke up on a Saturday with my best tool gone and no say in it.
That is not a failure of my system. It is the edge of what any system can do.
Own your architecture. It is still the most important thing you can do, and it is what let me keep working when the floor moved. But do not let anyone sell you the fantasy that it makes you untouchable. It does not. Nothing does.
It made me survivable. In a week like this one, that turned out to be the most any of us could ask for.
