There is a tool of immeasurable, unpredictable, and pervasive power. Those who master it gain an insurmountable advantage, yet taking control of it is no simple matter. It enables works of art, scientific discoveries, even spiritual revelations; at the same time, it also fuels rhetoric, propaganda, bureaucracy, falsehoods, and totalitarian regimes. Is it prudent to make it widely available? Can the masses handle it without harming themselves? Would it be better to leave it in the hands of small groups of competent people?
If you think I am talking about artificial intelligence, you are mistaken, because I have allowed myself a leap into the past. I am talking about writing, a tool that for a very long time remained the preserve of closed castes that controlled its spread in order to cultivate and preserve their own privileges.
At a certain point, through a decision that was anything but obvious and that faced prolonged opposition, it was concluded that teaching the masses to read and write was preferable to entrusting a small caste of chosen few with a monopoly over this technology. I believe that was the right decision, despite the harm that writing has caused as a result of becoming widespread. And I suspect that artificial intelligence will confront us with a similar crossroads.
Let us return to the present. Within the space of a few weeks, Anthropic announced Mythos while refusing to release it to the public, paired it with a “domesticated” version called Fable and made that publicly available, only to see both models banned by the United States government, which ordered their suspension for “any foreign national,” including Anthropic’s own non-US employees. A few days later, the restrictions were relaxed and Fable became accessible again. In the meantime, OpenAI unveiled GPT-5.6 and restricted its release to a handful of trusted partners, once again at Washington’s request.
The comic aspect of this story is the way Anthropic, and not Anthropic alone, has dug its own grave. After years of marketing and hype from tech companies about the extraordinary dangers posed by their tools, and if that does not seem ridiculous now, consider how ridiculous it sounded four years ago, the publicity campaign designed to attract investors appears to be turning against the very companies behind it. Cybersecurity researcher Peter Girnus explained the problem well: if you describe your product as “a munition” in every press release, sooner or later some government will take you at your word.
The technical pretext used to justify the shutdown is extremely weak. According to Anthropic’s own account, the government acted because of a filter bypass that consisted, in essence, of asking the model to read a piece of code and identify its flaws. Many other publicly available models, including OpenAI’s GPT-5.5, provide this capability without difficulty, and engineers use it every day. More than one hundred experts, including figures from Adobe and Nvidia, wrote to the authorities arguing that the Mythos models are rather good at finding vulnerabilities and that withholding such a tool from the public while competitors continue developing their own systems amounts to an own goal.
These are indeed highly capable models, but they resemble useful cognitive tools far more than weapons of mass destruction. Admittedly, we do not know whether a future version of Mythos might represent such a leap forward that it finally makes the tech companies’ advertising claims come true, whether it will be able to uncover previously unknown vulnerabilities in every operating system, or do something horrific that I currently struggle to imagine. Anthropic claims that the present version already identifies critical flaws in every browser and that, during a test conducted with intelligence agencies, it managed to uncover vulnerabilities in sensitive systems within a few hours. Perhaps these figures have been inflated for effect, but I would not bet on the threat remaining modest forever.
The question that interests me, however, is a different one: assuming that the tool becomes even more formidable, is keeping it private really the wisest choice? I doubt it.
When a tool of this kind remains in the hands of a small number of people, whether a government or a company, abuse by those who control it causes harm that is less visible and more difficult to counter than more widespread malicious use. Broad availability creates the risk of offensive applications, but it also enables distributed countermeasures and balances of power. Concentration produces nothing but asymmetry, and places that asymmetry in the hands of actors whom we have no reason whatsoever to consider trustworthy.
In a Facebook post, Giorgio Gilestro described how, for much of the Cold War, the United States treated cryptography as a weapon. In legal terms, exporting an algorithm with a strong encryption key was equivalent to exporting military equipment. Things changed in 1991, when Phil Zimmermann wrote PGP and released it as open-source software, becoming the target of a criminal investigation for the illegal export of weapons. The case collapsed against an irreconcilable contradiction: even if code is a weapon, printing it is an act of free expression, which in the United States is protected by the First Amendment.
Gilestro draws a conclusion that I share: today’s open weights are the PGP of our time, numbers released openly, reproducible by a student in any country, and impossible to retrieve once they have been distributed.
A Citigroup note reports a surge in demand for open-weight models as regulators restrict access to frontier systems, partly because the performance gap narrowed following the release of China’s GLM-5.2. This freely downloadable model cannot be blocked in the same way as Fable or Mythos and, more importantly, costs roughly one-tenth as much as its Western competitors. In other words, restrictions are driving a migration toward open, unrestricted, and inexpensive systems.
Companies such as Airbnb and Cursor have said that they are building part of their infrastructure on Qwen and Kimi, the open models developed by Alibaba and Moonshot, achieving performance comparable to Western systems at a fraction of the cost. On OpenRouter, a platform that routes requests among hundreds of different models, Chinese open-weight models have gone from a negligible share to approximately 61 percent of all tokens processed in the space of a year and a half.
Some American lawmakers have opened an investigation into Airbnb and Cursor precisely because of their use of those models, and Airbnb’s CEO was forced to clarify publicly that no sensitive data was being sent to Chinese developers. In truth, open models are the only systems in which data can remain genuinely private, provided they are run on a company’s own servers, regardless of the country in which the model originated.
Meanwhile, European officials have also expressed frustration at their dependence on decisions made in Washington. It is no coincidence that Europe has been entertaining the idea of bringing Anthropic into the fold and relocating it within the European Union.
But why are American companies not responding by cutting prices? To some extent, they already are. OpenAI is reportedly considering reductions in token prices, Microsoft has launched its own range of inexpensive models, and Google has introduced a lighter version of its systems while its CEO acknowledges that many companies had already exceeded their annual budgets by May. Even Nvidia has begun releasing its own models for free download, presenting them as an alternative both to Chinese systems and to closed American laboratories.
Cutting prices or offering smaller models, however, is not the same as releasing model weights. Even if Anthropic reduced its prices to Zhipu’s level, you would still remain the customer of a service that could be blocked from one day to the next by government decree, as happened with Mythos.
The problem is that these companies survive on investment, and investors may eventually lose patience. I am not an economist, but concluding that this will necessarily lead to the collapse of US companies or the bursting of the famous bubble, however real that bubble may be, strikes me as an oversimplification. There are still plenty of investors. Much will depend on the strategy adopted by the big technology companies, but it seems plausible that possessing the most powerful models will no longer be enough now that these systems have become infrastructure. Competitive prices will also be necessary, which means investing in reducing their energy and computing requirements.
It is nevertheless plausible to imagine that, sooner or later, open systems will catch up with the various Mythos models, rendering any prohibitionist strategy pointless. Perhaps one day we will retrace the path we followed with the alphabet, setting aside fear and privilege in order to make this cognitive tool as widely accessible as possible.
But I doubt that will happen tomorrow.
Francesco D’Isa