You get a call from an engineer in the early morning because the production system is down. Well, it's not all down, but the services that relied on Claude Fable 5 are not functioning. It took your team a few hours to reach this conclusion. In that time, customers are blowing up your phone and inbox wanting to know what is going on. Your engineers opened a browser and their development tools to try to use the Fable model directly only to see it wasn't responding. However, there were no announcements yet from Anthropic. You assume it is an outage with that model. You need to stop the bleed now. So, you change every API call that you make to the AI tooling to use a different model. It works, and the phone stops ringing.
A couple days later, a customer calls you. They noticed that the system seems to be behaving in odd ways. The customer can provide their data to the system, but the summarizing feature they rely on is not performing like it used to. You tell them you'll file a bug ticket and have the team look into it. Next week, you get another one of these calls. Except this customer is a lot angrier. They relied on the data your intelligence platform gave them, and it was inaccurate. Ultimately, they left their sales meeting embarrassed and lost the deal. That was the last straw for them, and this customer is taking their business somewhere else. Naturally, this escalates the bug ticket and you get involved because it needs to be solved today.
Your engineers reveal that nothing is technically broken. The services are running, they are logging interactions, and the data is being processed as expected. No errors, no exceptions, no corruption. The only thing that changed was the LLM you are using to process customer data. Since the incident, Anthropic issued a notice that Fable is no longer available due to government intervention. The model your competitive advantage was built on is gone, and there is nothing you can do about it. The next best model is simply not good enough. There is no way you can replicate Fable on your own. You know that one customer complaint, and one customer leaving is just the start unless you can figure out a solution. But you know there isn't a good one, and likely none of them are quick.
Panic starts to set in. Is this the end of the business? Is this the end of your tenure here? You bet everything on having this model available, and pushed building a backup plan off to the future. The weight of this is crushing and your team is patiently waiting for direction. The last thing you want to do is pass that panic on. So, you ask your best engineers to assess the models you have access to. Then you quietly sit with the weight of it. You made a big mistake and now you have to face the consequences.
Had the Fable model been available for longer, this story wouldn't be a fable. It would be someone's reality. The good news is that it wasn't out long enough for any companies to build their business model around it. But that same risk applies to every LLM you depend on.
Each LLM has its own personality, its own variance of outputs. Some of the models align perfectly with the type of output that makes your features shine. Other models may be useful, but users will know that something is off. That difference is what earns or loses customers.
A few days ago, Anthropic issued a precedent-setting notice that their Fable and Mythos models would no longer be available due to government restriction. It states that the U.S. government has identified, or at least suspects, potential vulnerabilities that pose a risk to national security. Whether this is the real situation or there are other reasons behind the restriction doesn't matter. It shows that these models can be taken away in an instant.
The appropriate response is to build a system that is resilient to this kind of vulnerability. Treat it like any other vulnerability. Consider the options, pick the best one, and start evolving your products. If one model goes down, you need to have a backup that still meets the users' needs and expectations. If you have to use an AI system to deliver a certain experience, then your system needs to be able to swap to another model when one doesn't respond. That way, the user always gets a response.
Another option would be to use an LLM that you have full control over. Whether that runs locally or remotely is completely up to you, and should be determined through analysis. The open source models are constantly improving and will support a variety of applications. You can even train a model for your specific domain and business value. And it never hurts to ask if standard programming logic will suffice. The right answer is likely a combination, and that needs to be determined.
The cost of not building this is not abstract. Whenever you build around any capability, you are at the mercy of it being available. For established companies, it can be an expensive lesson. For AI-native businesses, it might cost you everything.
Someone in another country can disrupt your entire business model. You know that now. That should be enough to motivate everyone in the industry to change.
The landscape is constantly changing, but that is not an excuse to be unprepared. Don't let your business become the next fable.