In early June, Anthropic shipped Claude Fable 5 - by most accounts the most capable AI model you could buy. Within days, the US government issued an emergency export-control directive, and Anthropic abruptly switched it off. Not just for foreign users — for everyone. Its bigger sibling, Mythos 5, went dark too.
No detailed public explanation. No timeline. Companies that had already moved work onto the new model got a shutdown notice measured in hours.
Two and a half weeks later, the government lifted the restrictions, and access came back this week. Good news. But if you run a business, the restoration isn't the lesson. The outage is.
This wasn't a crash or an outage in the normal sense. Servers were fine. The company was fine. The model was simply not allowed to exist in public for 18 days. There's no SLA on earth that covers "the government turned it off," and no status page that predicts it.
Here's the uncomfortable question: if the AI tool at the center of your operation disappeared tomorrow — banned, retired, acquired, repriced, or just changed beyond recognition — what actually happens to your week?
I got a rehearsal of this earlier this summer. I run an automated system that depends on a Claude model, and the specific model it used was retired out from under it in June — normal product lifecycle, nothing dramatic. It broke anyway. The fix wasn't picking a better model; it was accepting that every model is temporary. The system now checks its model is alive on every startup, and if it isn't, it walks down an ordered fallback list and sends me an alert telling me which backup it's running on. It has already used that fallback path once.
What to actually do, at any size:
Know what's under the hood. For each AI tool you rely on, find out which model it runs on and what happens when that model changes. If the vendor can't answer, that's an answer.
Prefer tools that let you switch models. Model-agnostic is quietly becoming the standard — even Microsoft is building for it (see Quick Hits). A tool welded to one model is a bet on that model's continued existence.
Write a two-line plan B for anything critical. "If [tool] is down two weeks, we do [alternative], and [person] makes the call." That sentence costs nothing today and everything when you need it and don't have it.
Don't build load-bearing processes on week-one releases. Fable 5 was days old when it vanished. Let new models prove they'll still be there next month before they become the thing your business stands on.
Who this is for: anyone who has moved a real piece of their operation — support, quotes, scheduling, reporting — onto an AI tool. That's the right move. Just make it the way you'd hire: assume any single employee might quit without notice, and make sure the work survives it.
Quick Hits
Your website may be about to go invisible to AI — by default. Cloudflare, which sits in front of millions of websites, announced that starting September 15 it will block "mixed-use" AI crawlers by default, and it's evolving its Pay Per Crawl program into Pay Per Use, where AI companies pay sites for content they use. Why it matters: more and more customers ask ChatGPT or Claude for recommendations before they ever hit Google. If your site is behind Cloudflare, whether AI assistants can read it is about to become a setting — one you should choose on purpose, not inherit. Being citable in AI answers is turning into the new local SEO.
Even Microsoft won't marry one model anymore. Microsoft launched a $2.5 billion unit that puts roughly 6,000 of its engineers directly into customer companies — and the design detail that matters: it's built to let those companies swap AI models freely instead of standardizing on one. Trade press called it the end of the single-model era. Why it matters: the company most publicly wedded to a single AI partner is now engineering for switchability. That's the same posture this whole issue is about, at a $2.5B scale. Match it at yours.
Prompt of the Week: The Single-Point-of-Failure Audit
This week's theme is dependency, so here's the 20-minute exercise version. Paste this:
Act as a business continuity consultant. Interview me one question at a
time to map the single points of failure in my business — the tools,
vendors, people, and accounts that would stop work if they disappeared
tomorrow.
Cover:
- Software and AI tools: what breaks if each is unavailable for two weeks?
- The one person (including me) whose absence stalls everything
- Accounts or passwords only one person can access
- Data that exists in only one place
For each single point of failure, rate the impact 1-5, then give me the
cheapest realistic backup: an alternative tool, a documented workaround,
a shared-access process, or a copy of the data.
Finish with a one-page "If X disappears, we do Y" plan I can save.You'll finish with a short list of things that would genuinely hurt, and a cheap fix for each. Most owners find at least one five-out-of-five they'd never written down. Better to meet it in a doc than on a Tuesday morning.
Like what you're reading? Forward it to someone who'd get value from it. And if you're curious what AI could actually do inside your business, book a free 15-minute audit — no pitch, just a look at where you're leaving time on the table.