Are we really going to miss Mythos and Fable?

The United States has unexpectedly imposed export restrictions on Anthropic’s latest AI models: Claude Fable 5 and Claude Mythos 5. European companies can no longer use these models effective immediately. The immediate economic impact of this will be limited; the models had only just become available. There are likely no European companies yet that have built their entire business model around these specific systems.

Nevertheless, this is an unprecedented warning. Software as a Service is ephemeral. It is a service that can be terminated at the stroke of a pen by a foreign government. What exactly are these models, how significant is the strategic damage to Europe, and how should we proceed from here?

What are Mythos and Fable

If we strip away all the marketing hype, Claude Fable 5 and Claude Mythos 5 are simply Anthropic’s latest, extremely powerful language models (LLMs). They are better than their predecessors, but not fundamentally different in their basic architecture. The mythical naming and the sudden legal disputes suggest that something completely new and dangerous has been launched. It’s not quite that dramatic, though the focus of these models has indeed shifted.

In fact, it’s the same underlying model. Mythos 5 is exceptionally good at discovering weaknesses in software and reasons in a way that aligns perfectly with both cyber defense and offensive hacking. Because this entails significant risks, Mythos has always been available exclusively to a select group of U.S.-approved customers. Fable 5 was the public, widely accessible version: essentially the same model, but equipped with additional built-in security safeguards to prevent it from being misused for hacking or sabotage. That is, until the U.S. government intervened, fearing that these safeguards could be circumvented by malicious actors.

How did this model become so good?

The big question is how Mythos has become so effective at solving complex problems. Older models were primarily trained on massive amounts of “finished” text and code. They’ve seen billions of lines of perfectly working code, have read the textbooks, and know the FAQs. But they have little understanding of the “messy” process of development and debugging. After all, the frustrations, the trial-and-error, and the intermediate steps aren’t found in a polished codebase. Older models are therefore “book smart.” Mythos is “street smart.”

According to analysts, this “street smart” intelligence doesn’t come from the AI secretly eavesdropping on programmers’ screens or keyboards like some kind of spyware. But how did the model actually perceive human interaction?

We have to rely on external analyses; Anthropic considers the training method a trade secret. The secret lies in the long-term iterations within AI chat interfaces and development environments. The training data presumably consisted of millions of anonymized chat sessions, prompt chains, and logs from Anthropic’s own tools and APIs, sourced from business users and developers who explicitly shared their data for this purpose.

What the model saw in this data was not the live code editor, but the process of problem-solving. A programmer pastes a piece of malfunctioning code with an error message into the chat, the AI offers a suggestion, the programmer tries it out and replies: “No, that solution doesn’t work; now I’m getting a NullPointer exception,” after which the human and machine take a new turn together.

By analyzing precisely these complete chains of trial and error, the analysts argue, the model learned how a debugging process unfolds. It saw the raw error messages, the dead ends, and the human logic required to ultimately correct an error. So the model wasn’t looking over the user’s shoulder while they typed, but learned from the extensive iterations, frustrations, and corrections in the chat logs. Anthropic has permission from its customers—perhaps even from you—to use this data.

However, this creates a crucial advantage for American tech giants. To replicate this “street-smart” training, you need enormous amounts of data from the human work process. The parties that own this data are companies like Microsoft (owner of GitHub and VS Code), Google, and OpenAI. They provide the development tools used by programmers worldwide and thus capture the telemetry and behavior of developers. They will undoubtedly come up with similar products, driven by data for which we in Europe simply do not have the platforms.

It’s the data, stupid!

This should give us pause for thought in Europe. Human work, decision-making, mistakes, and the iterative process are the next frontier for training AI. The Fable and Mythos cases are a stark reminder to keep European data within Europe. We must ask ourselves whether we should blindly let Microsoft Teams take the minutes of our strategic meetings, and whether we should let American AI assistants monitor our software development.

If we want to take a stand in the EU, we must start collecting data ourselves and training AI models locally. In doing so, we immediately run up against our own regulations, such as the GDPR. In Europe, data is, in principle, the property of the person it pertains to, or of the creator, and not automatically of the tech company that stores it on its servers.

This is right and proper and does not constitute a major barrier to innovation, but it does require careful consideration. The GDPR does not prohibit training AI; it merely requires that this be done with data to which you have a legitimate right. You can easily make agreements with developers and companies, use anonymized or synthetic data, and transparently ask people for consent or compensation. Anything is allowed, as long as you set it up properly.

The major challenge, therefore, is not collecting the data in a legally correct manner. The challenge for Europe is that we don’t have that data. As long as everyone is using American tools en masse, we’re giving away valuable data for a song, only to buy it back later—if the U.S. government allows it—for a high token price. So it’s high time for European platforms that keep the data within our borders and learn from it locally.

The bigger lesson

Today’s discussion centers on Fable 5, a relatively new model that hasn’t yet been deeply embedded in business operations anywhere. But this geopolitical intervention exposes a massive vulnerability. What if the U.S. government were to shut down Google Gemini for European users? Or temporarily restrict Office 365 due to a trade conflict or security risk?

That may seem unlikely right now, but this case proves that we in Europe do not have control over our digital toolkit. As long as we depend on cloud services and AI models that fall under the jurisdiction and whims of another superpower, we are building our digital economy on borrowed ground.