Claude Fable 5 Vanished for 3 Weeks: 3 Moves Before You Trust AI Tools With a Deadline

Full transcript

Lightly edited for readability. The audio is the original.

0:00Claude Fable 5 vanished for everyone

On June 9th, Anthropic shipped the most powerful AI model the public had ever been allowed to touch. 3 days later, the US government made them pull it. And because of how the order was written, they couldn't just pull it overseas. They pulled it for everyone. You, me, every user on the planet.

Gone. Not throttled. Not limited to a few countries. Off. It stayed gone for a couple weeks. Then it came back like nothing happened.

That whole round trip, launch to gone to back, took less time than you get between 2 month-end closes. If you're a finance pro putting any part of your job on top of AI, that one month tells you more than any product demo will. Here's what that speed should change about how you work.

Welcome to Excel at AI in Finance, the show for finance pros who want to use AI to get home sooner and be the one their company can't afford to lose. I'm Mike, I run F9 Finance, and I'm a senior finance leader working in the trenches.

1:09A finance pro's AI close, mid-deadline

I want to tell you about someone in the Finance AI Lab. I asked her if I could share this on the show, and she said yes, as long as I changed her name and kept her company out of it. So let's call her Rachel.

Rachel's a senior FP&A analyst at a mid-sized company: manufacturing, a few hundred people. And over the last year inside the Lab, she built herself a real workflow. Every month, she takes her raw variance data, the actuals against budget, and she runs it through a set of prompts she built that turns it into a first draft of her commentary and the narrative for her reforecast. The stuff that used to eat a full day of writing, she had it down to about an hour.

Now, Rachel has one habit that matters for this story. She always runs that workflow on the newest, most powerful model she can get her hands on. She's not precious about which one. Whatever's best that month, that's what she uses.

So when Fable 5 dropped on June 9th, and everybody said it was a step up, she moved her whole close workflow onto it that same week. The run was great. Cleaner first drafts, less cleanup. She was thrilled.

And I want to be clear. This is a person who does everything right. She's careful. She checks her outputs. She's not reckless with AI at all.

3 days later, she sits down mid-close to run her main numbers. Board pack is due Friday. She opens it up, and the model is just gone. Not slow, not throwing an error she recognized. Gone. And the message she's getting doesn't say come back in an hour. It doesn't say anything useful at all.

She refreshes it, she checks if it's just her, she pops into a forum, and there it is. It's not just her, and it's not her company. The model is off for everybody, and nobody's saying when it's coming back.

Now, here's the part I want you to sit with. A year ago, that moment would have wrecked her week. She'd be staring at a blank commentary the day before the board pack was due, with the one tool that made it fast suddenly missing.

But that's not where Rachel was anymore. Because the thing she'd built went deeper than a button that says "do my variance commentary." She understood what the workflow was doing, step by step, and why.

So she did the thing a year of practice made automatic. She pulled the workflow over to a different model that afternoon. She adjusted a few prompts, because the new model needed slightly different instructions, and she shipped the board pack on time. Her CFO never even knew there'd been a fire.

From the outside, it was a normal close. On the inside, the most powerful tool she owned had vanished with 3 days' notice and no explanation. And it didn't cost her the deadline.

Now, Rachel got lucky in one important way. She'd built the skill underneath the tool. So when the tool moved, she moved with it. A lot of finance teams found out that same week that they'd only built the shortcut. And that gap between the people who could roll with it and the people who got stuck is the whole reason I wanted to do this episode.

4:01What actually happened to Fable 5

So let me back up and walk you through exactly what happened with Fable 5, and then what it should change for you. So here's the whole thing, start to finish. You don't need to be a machine learning person to get why this one matters to your job.

On June 9th, Anthropic released Claude Fable 5, their most capable model ever at the time. This is the tier that can work on a problem for hours on its own. It plans across steps. It writes its own checks. It catches its own mistakes. For finance, that's the level where the model stops finishing your sentence and starts doing the whole analysis.

Anthropic said this class of model can run for days on a single task. For the kind of work that used to mean a late night and 3 cups of coffee, that's a huge jump. It was priced like a premium tool, and it still flew off the shelf, because for a lot of knowledge work it was the biggest single jump anyone had felt in a while. People moved real workflows onto it fast. Rachel was one of thousands.

3 days later, on June 12th, the Commerce Secretary sent Anthropic's CEO a letter. It put export controls on Fable 5 and its bigger sibling, Mythos. And here's the wording that made it messy. It barred access by any foreign national, whether they were inside or outside the United States.

Think about what that means for a company running a public AI product. They have no way to check the nationality of every person typing into the thing in real time. So faced with an order they had to follow that day, Anthropic did the only thing they could. Within hours, they switched both models off for everyone, not just users overseas. Every single user.

Now, why did any of this happen? Well, a research team at Amazon had found a way to jailbreak Fable 5. They got it to identify, and in one case write working code to exploit, a software vulnerability. A live security hole. That's what put it in export control territory, the same bucket as technology that can be turned into a weapon.

And then it just sat there dark for about 19 days. No frontier model of that class available to anybody for almost 3 weeks. On June 30th, the Commerce Department lifted the controls. July 1st, Fable 5 came back worldwide, this time with a new safety filter built specifically to catch that exact jailbreak. It blocks it more than 99% of the time now.

The comeback itself was telling. They rebuilt the safety around it first, doubled the team working on it, and shipped it back harder to break. Even the fix came fast. So the full arc, ship it, cut the access, bring it back, fit inside about 3 weeks.

Analysts were already calling it the biggest AI governance story of the year. And I bet most of the finance teams using AI every day barely registered it. Here's why I care about this as a finance person, and it's got almost nothing to do with this security headline.

7:35Why this matters for AI in finance

Think about the tools you depend on to do your job. Excel doesn't disappear on a Tuesday. Your ERP doesn't get pulled by a government order over a weekend. The stuff finance runs on is boring on purpose. And boring in finance is a feature. Boring means it's there every single month when you need it.

That reliability is invisible until it's gone. Nobody thanks Excel for showing up, but the whole reason you can build a close around it is that it's always there.

Frontier AI is a different kind of dependency, and Fable 5 just showed everybody how different. The single most capable model on the planet went completely unavailable for everyone for 3 weeks, for reasons that had nothing to do with you, your company, your data, or whether your invoice was paid. A security report, a government letter, and it was gone.

Now play that forward into a finance function. Let's say you'd handed that model your 13-week cash flow build, or the first draft of every variance story, or your board narrative. Say it quickly became the thing that let you leave at a reasonable hour. And then, on a random Tuesday, it's gone, for reasons you can't influence and can't appeal. That stopped being hypothetical in June.

8:48Model deprecation: the new continuity risk

Here's the number that reframed the whole thing for me. When an AI model normally gets retired, the provider gives you notice. 3 to 6 months, usually, plus a migration path to the next one. That's the civilized version.

Fable 5 got 0. Gone in hours, no planning window, no wind-down. One analyst put it exactly right: a suspension like this strips out live operating capability with no planning period at all. This is a very different animal from your typical software sunset.

And notice who got a say in it. Not the finance teams who built on it. Not even the company that made it. A letter landed, and the most capable tool in the world went off the same day. When your dependency can be switched off by somebody 3 steps removed from you, well, that dependency is worth a second look.

And most teams made it worse without knowing it. The risk people are starting to name is wiring one specific model straight into a mission-critical process with nothing in between. No fallback, no layer that lets you switch. When you do that and the model changes or disappears, the whole process breaks. And it breaks with far less warning than any old system ever gave you.

Rachel dodged that because she understood her workflow well enough to move it. The teams that got stuck had bolted themselves to one model with no way out.

And before you write this off as a one-time freak event, look at the wider pattern. This same year, OpenAI is retiring dozens of its own models, one of the biggest deprecation waves the industry has seen. Google, Amazon, all of them are cycling models in and out on a schedule that keeps getting faster.

There's a phrase showing up in enterprise risk circles for this now. They're calling model deprecation the new continuity risk. And it's landing at the board level, because a model you built a critical process on can change or vanish with far less warning than anything finance has ever run on before.

10:37Concentration risk in disguise

And look, you already know this risk. You just know it by a different name. It's concentration risk. It's the same instinct that makes you nervous when 40% of revenue rides on one customer, or when one person on the team is the only one who understands the model.

We manage that kind of single point of failure all day long in finance. We diversify it. We document it. We build a backup. Somehow, when it came to AI tools, a lot of very careful finance people forgot their own playbook.

The teams that got burned in June had just never started thinking of their AI model as a vendor, one with a cancellation clause they didn't control. That's the real story under the Fable 5 headline. The one model and the one government order are almost beside the point. What matters is what it means to build your work on a layer that moves this fast.

So what did everybody take away from this? Well, the coverage was loud, and most of it, in my read, missed the question that matters most. There were basically 2 camps. Camp 1 said, see, this proves AI is too dangerous. A model that can write exploit code shouldn't be out there at all.

Camp 2 said the opposite: that this was government overreach, that clumsy regulation is going to hand the lead to Chinese labs while we tie our own hands. Some investors were genuinely angry about that second point, arguing that every day Fable 5 sat dark, well, the competition got a free head start.

And then there was a third thread from the governance people, pointing out that Washington still has no real process for this. When the government wants to move fast on a frontier model, it improvises every time.

And watching it play out, all 3 camps sounded right, and all 3 sounded like they were describing somebody else's problem. The safety people were arguing about the model. The market people were arguing about China. The policy people were arguing about process. Not one of them was standing where you're standing, with a deadline and a tool that just disappeared.

Because if you do real finance work, you don't get to vote the frontier off the island. And you don't get to wave the regulation away either. You just have to operate in a world where the best tool you have is also the least stable one.

So the read that matters sits underneath all 3 camps, and it's this. That volatility you saw with Fable 5 is not a freak event. It's a standing feature of working at the edge of AI now, and it's going to keep happening.

And here's the trap. The natural finance reaction to all of this is to pull back, to say, you know what? This stuff is too unsettled. I'll wait until it calms down. But that is the move that costs you the most, because the pace of AI is the real thing coming for your seat. And the finance pros who already decided to wait it out are the ones falling behind the fastest. The churn is the exact thing you need to get good at living with.

13:233 ways to protect your finance work

So let me make this useful. 3 things you can do this month, and none of them require you to be a technologist.

Number 1: for anything with a hard deadline, your close, your board deck, your reforecast, keep a backup path. A second model you know how to use, or a manual version you could fall back to. The rule is simple. Never let one model be the only thing standing between you and a due date.

Rachel made her deadline because she had somewhere to go. It doesn't have to be fancy. Even knowing which second tool you'd reach for, and having tried it once, is most of the game.

2: split your experimenting from your production. Go play with the newest, most powerful model the day it drops. That's how you stay sharp. But your recurring, deadline-bound work should run on something stable you can swap out without a fire drill. Chase the frontier on your own time, not on the board's timeline.

3, and this is the big one: invest in the skill, not the tool. This is exactly what saved Rachel. The prompt patterns, the workflows, the judgment for what to double-check before it goes to the CFO, all of that carries from one model to the next.

Practically, that means when you build something in AI, spend the extra 10 minutes to understand why it works, not just that it works. Save your prompts where you can find them. Write down the steps. The specific model won't be the newest one for long. What you know how to do with it stays yours.

Now there's even a name forming for this discipline inside bigger companies. They're calling it AI continuity management. Most teams haven't staffed it or even thought about it. You can be the person on your team who already has.

If you want the 15-second version of all 3, here it is. Stay close enough to AI that you stay fluent. Keep your critical work on ground you control. And put your bet on what you can carry from one tool to the next, because that's the part nobody can pull from you.

And that goes way past this one model. The finance pros who learn AI don't just keep their jobs. They become the ones their companies can't afford to lose. And a huge part of the reason is exactly what we're talking about here. When the tools churn, and they are going to keep churning, the person who can pick up the next one without breaking stride is the person everyone else in the building turns to.

This is the pace now. The most powerful tool in the world can vanish over a weekend and come back before you finish a close. The finance pro who can keep up with that pace is the one their company can't afford to lose.

15:49Counterpoint: when it didn't touch you

Now, before this gets too tidy, let me push back on my own take. And here's the honest counter. For a lot of you listening, Fable 5 going dark for 3 weeks changed nothing about your day. You weren't running the raw frontier model in production. You were in Excel with Copilot, or inside a vendor tool running on an old approved model, and none of that blinked. If you're on the boring enterprise stack, this whole saga was a spectator sport.

And there's a second thing I skipped past. You could argue the system worked exactly the way it's supposed to. A serious vulnerability got caught, the model got pulled, it got patched, and it came back safer. That's the guardrails doing their job.

So don't take this episode as "AI is too shaky, wait till it settles down." That's one conclusion that would cost you the most.

Now, here's the tell for which side of this you're on. If a model going offline for 3 weeks would have blown up one of your deadlines, you're leaning too hard on a single tool, and that's worth fixing immediately. If it wouldn't have touched you at all, your risk runs the other way. You're not close enough to the frontier yet to be building the skill that keeps you ahead.

Either way, the fix is the same. Get good enough that no single tool owns you.