# The war for intelligence has already begun

## Frontier models now have borders

**In less than a month, the American government switched off Anthropic's most advanced models and conditioned the launch of GPT-5.6. The builders themselves are calling for brakes they know they cannot pull. China advances in silence. And Europe applauds national models without understanding what the frontier is.**

*By Pedro Seabra*

On the night of 12 June 2026, at 5:21pm Washington time, a letter changed the history of artificial intelligence.

Three days earlier, Anthropic had launched Claude Fable 5, the most capable model ever made available to the public, the first representative of a new class, Mythos, which the company itself had described months before as too powerful in the cybersecurity domain to be released freely. Mythos 5, the unrestricted version, was reserved for a vetted group of organisations through the Glasswing programme.

The letter came from Commerce Secretary Howard Lutnick. Invoking national security and export control powers, it ordered the immediate suspension of access to Fable 5 and Mythos 5 by any foreign national, inside or outside the United States, including Anthropic's own foreign employees.

The company had no way of verifying, in real time, the nationality of millions of users. To comply with the order, it switched off both models for everyone. The entire world lost access because Washington wanted to exclude foreigners.

A commercial model, used by hundreds of millions of people, taken off the air by government decision, three days after launch.

It was not an isolated episode. It was the first chapter of a month that should have woken Europe up. It did not.

## Two cases, one new regime

What happened to Anthropic and what happened to OpenAI in the following weeks are two faces of the same phenomenon.

In Anthropic's case, the origin appears to have been a report about an alleged jailbreak, a technique for bypassing the model's safeguards, which researchers at Amazon, simultaneously an investor in and cloud provider to the company, are said to have taken directly to administration officials. At least, that is the story that has been made public. Anthropic publicly contested the severity of the finding, describing it as a narrow, non-universal jailbreak that merely surfaced minor, already known vulnerabilities, within reach of other models on the market, including OpenAI's GPT-5.5, which was not subjected to any restriction.

The company complied with the order but put its disagreement in writing: if this standard were applied across the industry, it would essentially halt all new deployments by all frontier labs. And it added a principle worth underlining: the government should be able to block unsafe deployments, but through a statutory process that is transparent, fair, clear, and grounded in technical facts. This action, the company said, did not adhere to those principles.

For 18 days, the models remained inaccessible. Anthropic sent negotiators to Washington. According to the American press, co-founder Tom Brown replaced CEO Dario Amodei as the main interlocutor, a sign of how personal the conflict had become between the company and an administration that had already banned federal agencies from using its models and classified Anthropic as a supply chain risk for the Pentagon.

On 30 June, the Commerce Department lifted the controls. Fable 5 returned the following day, now with an additional safety classifier trained in collaboration with the government. Mythos 5 remains restricted to roughly a hundred vetted American organisations tied to critical infrastructure and cybersecurity.

OpenAI's case followed almost without pause, but with a different choreography.

The company presented GPT-5.6 on 26 June and, at the administration's request, limited access to around 20 approved partners whose identities were shared with the authorities. For weeks, the Commerce Department's Center for AI Standards and Innovation evaluated the model, with OpenAI specialists stationed in Washington to answer the reviewers' questions. The system's capabilities in coding, biology and cybersecurity were at the centre of the scrutiny.

Only on 9 July, with a green light from the administration, did the GPT-5.6 family, christened Sol, Terra and Luna, reach the general public. Sam Altman described the process as a collaborative back and forth with the secretaries of Commerce and the Treasury and the national cyber director. But, according to press reports, he told his own employees that a government-curated access list is not a sustainable model.

Notice what these two stories have in common and what separates them.

In common: for the first time, the timing and conditions for launching the world's most advanced AI systems passed through Washington. No public law requires it. There is only an executive order signed by Donald Trump on 2 June, which created a nominally voluntary mechanism for the prior submission of frontier models to the government. In practice, voluntary became mandatory.

In the difference lies the lesson. OpenAI, which cultivated its relationship with the administration, negotiated a staggered launch and came out with its product on the market and mutual praise. Anthropic, on a collision course with the White House for refusing to accept unrestricted military uses, saw its models switched off by decree. The technology was comparable. The treatment was not.

When access to frontier intelligence depends on the political relationship between a company and a government, we have left the territory of products and entered the territory of power.

## The builders are calling for brakes they cannot pull

The most disturbing thing about this moment is not state intervention. It is what the creators of the technology themselves are saying.

On 4 June, days before the launch of Fable 5, Anthropic published a text entitled "When AI Builds Itself" that went almost unnoticed outside the industry and that deserved to be on the front page of every European newspaper. In it, the company revealed that more than 80% of the code merged into its own systems is now written by Claude, that its engineers produce roughly eight times as much code per quarter as they did before 2025, and that the trajectory points towards recursive self-improvement: systems capable of designing, building and training their own successors.

And it concluded with an extraordinary sentence coming from one of the leaders of the race: it would be good for the world to have the option to slow or temporarily pause frontier AI development, so that societal structures and alignment research can keep up with the technology.

Confronted with his own appeal, Dario Amodei, Anthropic's CEO, explained the paradox on television: a pause would only work if all relevant players, including geopolitical adversaries, accepted it in a verifiable way. He does not believe China would be a responsible actor in that scenario. Even so, he insisted, "we should be willing to at least try".

Days later, on 10 June, Amodei published a policy essay, "Policy on the AI Exponential", in which he took the step the industry had avoided for years: he argued that the government should have the legal power to block or deter the deployment of models that present unacceptable risks, subject to third-party assessment, and that the era of mere transparency is over. What is needed, he wrote, is serious and binding regulation, with mandatory testing like that required of cars, aeroplanes or medicines. The problem, he summed up, is one of speed: AI advances at an exponential pace and legislation moves slowly.

There is something vertiginous about this position. The man whose company had just been switched off by the government is asking for more power for the government. Not the arbitrary power that struck him, but power with rules. He builds as fast as he can, because stopping alone would mean handing the race to those who do not share his concerns, and simultaneously asks that someone build the brake he himself cannot pull.

At the other end of the corporate spectrum, Alex Karp, Palantir's CEO, took to the CNBC stage on 1 July for one of the harshest attacks ever on the frontier labs. He accused them of "irresponsibly" overstating the dangerousness of their models without clarifying how they protect their customers' intellectual property, of overcharging for tokens that generate no value, and of absorbing the proprietary knowledge of the companies they serve. On the strategic dimension he was even more direct, asking whether America is really going to outsource the country's battlefield to the consensus of Silicon Valley.

Karp is not a neutral critic: Palantir sells precisely the sovereignty layer it presents as the solution, and had just announced with Nvidia a sovereign AI operating system for governments. But the diagnosis he delivered live is hard to ignore: American companies are migrating to cheaper Chinese models at the precise moment the Trump administration is blocking access to the best American models.

Amodei and Karp disagree on almost everything. One built his company on the thesis that AI is too dangerous to be left without rules; the other built his on the thesis that hesitation is the greatest threat of all. But they converge on one point: nobody can stop. Amodei because adversaries would not stop. Karp because stopping would mean losing. The race continues because no runner trusts the others to slow down at the same time.

It is the exact definition of a collective trap. And it is accelerating.

## When intelligence manufactures intelligence

The reason this trap is different from every previous technological race lies in the numbers Anthropic itself disclosed.

In one of the published indicators, Mythos Preview outperformed human judgement in 64% of decisions about the next step of an investigation, against 51% achieved by Opus 4.5 a few months earlier. Jack Clark, the company's head of policy, admitted that some models could be capable of recursive self-improvement within two years.

Rigour is called for. Today's models do not create themselves. They depend on human teams, data centres, energy, business decisions and supervision. And there are legitimate sceptics who see in these announcements, published days after Anthropic began its stock market listing process, as much marketing as science.

But it is also no longer accurate to describe these systems as passive tools. They write the training code, analyse results, propose experiments, find errors, generate synthetic data and guide research decisions. A growing share of the work needed to create the next generation is done by the previous generation. In April, a single Claude agent produced more than 800 fixes in one of Anthropic's own systems, work a human engineer estimated would take four years.

When a technology begins to accelerate its own development, the interval between a manageable capability and a dangerous one shrinks drastically. It is this compression of time that explains the anxiety of governments, the conditioned launches, the suspension orders. States have realised they are no longer regulating products. They are trying to catch a train that is accelerating.

## China chooses another weapon

While Washington experiments with ways to control the circulation of its models, China executes, in silence, an opposite and perhaps more intelligent strategy.

Instead of locking away its best systems, several Chinese labs are publishing models with open weights, permissive licences and radically lower costs. Z.ai's GLM, with a one-million-token context window and an MIT licence, has become one of the most cited examples; independent analysts such as Simon Willison have described it as probably the most powerful open model available. Moonshot's Kimi models and DeepSeek follow the same path.

The effect is no longer theoretical. Coinbase revealed it cut nearly 50% of its internal AI spending by migrating its engineers to Chinese open models. Microsoft is considering using a fine-tuned version of DeepSeek as a lower-cost engine for its Copilot. American companies, at the heart of the American ecosystem, building on Chinese intelligence, in the very month the American government was switching off American intelligence.

The irony is brutal and the lesson is strategic. A model that is good enough, cheap and downloadable can have more influence than a superior model whose access depends on a foreign political decision. Whoever downloads the weights and hosts them on their own servers gains what no closed American system can guarantee: nobody can switch it off on them.

Every integration creates technical dependency. Every adaptation trains specialists. Every compatible tool reinforces the ecosystem. And, in parallel, Beijing reduces year after year its dependence on Western chips, investing in its own semiconductor chain under the pressure of sanctions which, paradoxically, accelerated that push for autonomy.

None of this is liberal innocence. The Chinese industry operates under strong state supervision, and openness towards the outside coexists with tight control inside. But it is precisely this selectiveness that makes it effective. China opens what wins over developers, markets and standards. It closes what it considers critical. The United States offers the world the opposite argument: American technology is better, but it can be taken away from you tomorrow.

For much of the world, that difference will be decisive.

## The strange European silence

And Europe?

Let us be exact about what happened and about what did not. Fable 5 was three days old when it was switched off. No hospital stopped, no company saw critical processes collapse, because operationally there were not yet any systems dependent on that model. It was not a catastrophe: the suspension merely blocked access to the most powerful and most capable system on the market, the one the whole world had its eyes on.

It was worse: it was a dress rehearsal.

For 18 days, the most advanced system ever made available to the public was inaccessible by unilateral decision of a foreign government, and European citizens were treated, for legal purposes, exactly like citizens of strategic adversaries: foreigners to be excluded. The mechanism was demonstrated, tested and normalised. Next time, the model switched off may have two years of integrations, critical processes and accumulated dependencies in banks, hospitals, courts, public administrations and companies across the European continent.

The absence of immediate damage perhaps explains the absence of reaction. It does not excuse it. It was precisely now, before the dependencies harden, that the alarm should have sounded. Red flags exist to be seen before the fire, not during it.

There was no coordinated statement from Brussels. There was no demand for transparent criteria for the treatment of allies. There was no parliamentary debate of any scale about what it means to depend on systems that Washington can switch off at dinner time.

The silence has an uncomfortable explanation: it is easier to protest against a dependency when a credible alternative exists. Europe does not have one. Confronting Washington would mean publicly admitting its own fragility.

It would be unfair to say the European Union is standing still. There is the AI Continent Action Plan, there are the AI factories attached to supercomputers, there is the proposed Cloud and AI Development Act, there is a EuroHPC competition to fund a European frontier model project. These are real initiatives.

But the numbers give away the scale. The EuroHPC competition offers the winning project, the European frontier model candidate, up to 2.5% of the total capacity of the European network's public supercomputers, for one year. That is the ceiling of the continent's ambition: a fraction of shared infrastructure, for a limited time, awarded by tender. Across the Atlantic and in China, the major labs operate dedicated data centres, built exclusively to train their models, permanently expanding and fuelled by annual investments of tens of billions of dollars. It is not a difference of percentage. It is a difference of nature. Europe has excellent universities and top researchers. It does not have the concentration of capital, compute, energy and distribution needed to contest the frontier.

And the AI Act, necessary work though it is, solves none of this. A regulation can define the conditions under which an American model operates in Europe. It does not guarantee that the model will remain available. It can oblige a provider to document risks. It does not create the chips, the data centres or the researchers to build an alternative. Regulation is a component of sovereignty. It is not a substitute for it.

Europe has become expert at creating rights over infrastructure built by others. But there is a fundamental difference between influencing the behaviour of an infrastructure and controlling it. The first grants legal power. The second grants strategic capability.

## Amália and the illusion of big words

On 1 July, the very day Fable 5 came back online by Washington's decision, Portugal presented at the Técnico Innovation Center the Amália, the first open large language model developed for European Portuguese.

Prime Minister Luís Montenegro said the model equips the country to face the coming decades "with greater autonomy, sovereignty and less dependence". He argued that Portugal should be a promoter and beacon for the challenges of the future, celebrated the culture of risk, and acknowledged, with a frankness that deserves to be noted, the starting position: "we are dependent", he said, referring to the fact that the major developments come either from the United States or from China.

Let us be fair to the project before being hard on the discourse.

Amália answers a real problem. European Portuguese is under-represented in the large international datasets, and many models confuse our language with Brazilian Portuguese. Sixty researchers from public universities and research centres built, in 18 months, an open, auditable, multimodal model that can run on sovereign infrastructure, aimed at applications in public administration, education, culture and defence. It is useful, competent and necessary work. Portugal should have models adapted to its language and its institutions.

The problem begins when one measures the distance between the words and the capability.

Because the essential question is not budgetary. It is what a model can do.

What defines the frontier is not speaking a language well. It is the capacity for reasoning: working autonomously for hours on complex problems, programming entire systems, doing scientific research, analysing millions of documents, finding what no human would find in the time available. And, above all, it is the effect of that capability on those who use it. A frontier model is not a tool that answers: it is an amplifier that transforms the productivity of an engineer, the diagnostic capacity of a doctor, the research speed of a university, the competitiveness of an SME that suddenly does with five people what used to require fifty.

That is the true power of artificial intelligence: to augment human beings and organisations. The companies building today on Fable 5 or GPT-5.6 are not merely adopting technology. They are multiplying the capabilities of their people. And it is in that multiplication, not in the language of the output, that the global competitiveness of the coming decades will be decided.

Amália, with 9 billion parameters, does not play in that league, nor was it designed to: the project's own technical leads reject the comparison with generalist frontier systems. Speaking correct European Portuguese is necessary and solves a real problem of linguistic representation. But it is not what makes a Portuguese worker or company more capable against competitors who operate side by side with systems of another cognitive order of magnitude. A Portuguese professional limited to national models will be competing, within a few years, with foreign professionals amplified by frontier intelligence. That asymmetry, not syntax, is the real sovereignty question.

It is the political discourse that fails to say this with such clarity. Talking about sovereignty and self-determination in reference to a specialised national model, at the precise moment when the stroke of a pen in Washington demonstrates the power to switch off the systems on which the competitiveness of European companies and states will depend, reveals a dangerous misunderstanding of what frontier models are and of what is really at stake.

Because what is at stake is not having an assistant that speaks good Portuguese. It is knowing who controls the intelligence that, within a few years, will be embedded in tax administration, health systems, defence, banking, energy and the very democratic process of each nation. It is knowing whether a country's citizens and companies will have access to the capability that makes them competitive, or whether that capability will permanently depend on the political goodwill of another power. A country that does not grasp the difference between a specialised national model and frontier capability is not building sovereignty. It is building a sensation of sovereignty. And sensations do not withstand export control directives.

A national model does not create national sovereignty. A data centre does not create a sovereign cloud. A regulation does not create an industry. And a succession of inaugurations with speeches about autonomy does not constitute a strategy.

## Three possible Europes

Europe is approaching a choice it has not yet had the courage to own.

The first option is to accept American dependence: to keep using models from the United States, protect data, impose rules, negotiate terms. It is cheap and pragmatic. It leaves the continent exposed to decisions like that of 12 June, taken without warning and without recourse.

The second is to build sovereignty at the application layer: specialised models like Amália, industrial data, critical applications in health, energy, defence and public administration, accepting external base models where necessary. It is realistic and can generate significant value. It keeps the dependency at the core.

The third is to contest the frontier: tens of billions of euros, abundant energy, large-scale data centres, European public procurement, aggressive talent attraction, deep venture capital, a common policy for open models, and the acceptance of failures along the way. It would require treating AI the way the United States treated its space programme: a strategic capability, not a collection of projects with a ribbon to cut.

Europe seems to want the benefits of the third option while paying the price of the first and making the speeches of the second.

## The question that remains

June 2026 delivered the clearest warning Europe has ever received.

A government switched off by decree the most advanced models of one of its own companies, and Europe and the rest of the world were locked out, without a public word about what had happened. The same government conditioned the next launch by its biggest lab, choosing who could access it and when. The creators of the technology came out publicly to say that it is approaching the point where it builds itself, and that they would like to be able to slow down, but cannot, because nobody else would slow down. China answered all of this with open weights, low prices and a growing share inside American companies themselves.

And Portugal presented a national model with speeches about sovereignty, as if speaking the language were the same as possessing the intelligence.

For years, the European debate asked how to control artificial intelligence. The question of the next decade is a different one: what share of that intelligence does Europe actually control?

Because a region can have the best rules in the world and remain dependent on the decisions of another. It can possess rights without possessing capability. It can create national models and remain absent from the frontier. It can proclaim autonomy and discover, on some ordinary night at 5:21pm, that the system it depends on has been switched off and that no European alternative stands ready to take its place.

The war for intelligence has already begun. The United States is fighting it with decrees. China with open weights. The labs with appeals they cannot honour.

Europe, for now, is fighting it with press releases.