Anthropic Sets a Precedent: The Most Powerful Artificial Intelligence Will Not Be for Everyone

Yesterday, Anthropic, the American company developing the chatbot Claude, introduced two new artificial intelligence models, which are updated versions of the system that powers its digital assistants. They are called Claude Fable 5 and Claude Mythos 5 and share the same brain. In the past, launches of this kind were narrated through rankings of scores in tests. This time, the detail that deserves attention lies in how the two models are distributed, as for the first time a company openly states that the full version of its technology is too powerful to be made available to everyone.

Mythos 5 is the unrestricted version, reserved for a select group of approved partners, including cybersecurity operators and critical infrastructure managers, through a program called Project Glasswing, created in collaboration with the U.S. government. The reason is that some of its capabilities are dual-use: the same model that helps a company find vulnerabilities in its IT systems can assist a criminal in attacking them, and the same skills that accelerate pharmaceutical research can facilitate dangerous applications in biology.

Fable 5 is the twin destined for the public. It has the same capabilities, but when a request touches on sensitive topics, from cybersecurity to chemistry, automated systems intercept it and redirect it to an earlier, less advanced, and considered safer model. According to the company, this happens in less than five percent of work sessions, and during the rest of the time, the user is effectively using a system almost identical to that reserved for partners. So far, models were distributed based on price, and from today the distribution criteria include risk as well.

As for what these systems can do, the declared strength is their ability to work independently for long periods on complex tasks. During testing, Anthropic reports, Fable 5 completed in a single day for the payments company Stripe a software update that a group of programmers would have taken more than two months to finish.

Ethan Mollick, a professor at the Wharton School and one of the most listened-to observers in the field, had early access to the model and describes an experience that he calls both delightful and unsettling. He asked it to develop software for his academic research, a tool the scientific community had been waiting for years and that no one had ever developed because there was no market to cover the costs. The system produced a 19-page design document on its own and then worked for nine and a half hours without human intervention, recruiting smaller, cheaper copies of itself as assistants. In the end, it delivered a working program.

The word Mollick uses to describe his role in all this is "client." His job, he writes, has been reduced to describing the result he wants and evaluating what is delivered to him, while the hundreds of intermediate decisions are made in a process he cannot observe. The comparison he proposes is that of a production studio, where the client signs off on the final approval without ever having set foot in the departments.

Together, the two accounts suggest at least two consequences that concern everyone, even those who do not work with artificial intelligence. The first relates to competence. The software delivered to Mollick contained errors, and he found them because he has been in that field for decades. To judge the work of a machine requires the experience of someone who has been doing that job for a long time with their own hands, and if machines absorb precisely the grunt work in which that experience was formed, it will be necessary to understand where the professionals capable of evaluating them will grow tomorrow.

The second consequence concerns access. These models are expensive: Fable 5 costs double its predecessor and consumes computational resources at remarkable rates. If the more capable versions are distributed based on risk and expenditure levels, the competitive advantage of a company or a professional will also depend on the ability to access the right version. For Europe, which buys these technologies rather than producing them, the issue is anything but abstract. And to the question of who can use the complete model, and under what criteria, today only one private company responds in agreement with a foreign government.

For workers, in the meantime, the practical advice is less futuristic than it seems: learn to describe precisely what you want to achieve and recognize when a well-packaged job is wrong. These are ancient skills, and for a good while, they will remain the part of the job that no one will be able to commission to a machine.