On 12 June 2026, Anthropic announced that the US government had ordered it to suspend access to Fable 5 and Mythos 5 for any foreign national, “whether inside or outside the United States”, including Anthropic’s own foreign-national employees. The company said it had therefore disabled both models for all customers while it worked out how to comply. This was not an ordinary country-based export restriction. It was a citizenship-based restriction on access to some of the most capable AI systems yet released by an American company.
That decision looks crude, and it is. But it becomes less mysterious when read alongside Anthropic’s own numbers. In its Institute essay on recursive self-improvement, Anthropic says that by the second quarter of 2026 its typical engineer was merging eight times as much code per day as in 2024. The broad point is obvious: this is no longer merely a better autocomplete. It is the beginning of a multiplier.
The multiplier is what governments are frightened of. Fable 5 and Mythos 5 are not being treated merely as consumer software, or even as powerful enterprise tools. They are being treated as accelerators that may help capable users build the next generation of accelerators. If Anthropic is broadly right, then the danger is not simply that a foreign user might ask Fable 5 a dangerous question. The danger is that a foreign lab, company or state might use Fable 5 to shorten its path to the system after Fable 5.
The best way to grasp this is to think about time in war. Imagine Nazi Germany possessing a ×100 research multiplier in 1939. The point is not that Germany suddenly acquires Bletchley Park or Los Alamos. The point is that its own codebreaking, cryptographic security, radar work, jet aircraft development, rocket programme and nuclear research all begin to move through their dead ends much faster. Work that would have taken years of human trial and error could, at least on the cognitive side, be compressed into weeks or months.
That does not remove every physical bottleneck. Uranium still has to be mined and enriched. Reactors and factories still have to be built. Weapons still have to be tested. But strategic history is often decided before every bottleneck has disappeared. If Germany had broken Allied codes faster, protected its own communications better, avoided the wrong turns in nuclear research and identified the right industrial route to a bomb earlier, 1943 might not have been the year in which the Axis began visibly to lose the war. It might have been the year in which Washington was forced to capitulate.
That is the point. A huge multiplier does not merely save labour. It brings decisive possibilities forward in time. The weapon is not only the final artefact: the bomb, the broken cipher, the autonomous weapon, the next AI model. The weapon is acceleration.
This is why the AI race is not really a race to build the best chatbot. It is a race for the multiplier that can accelerate research, software development, cyber operations, model training, evaluation design and AI development itself. A model that helps write software is useful. A model that helps build better AI is something else. It is a tool for producing the next tool.
The politics of such a system depends on how large the multiplier becomes. A ×2 multiplier is productivity software. A government can allow that to spread widely. A ×5 multiplier is more sensitive: it might be shared with allies, trusted companies and regulated sectors. A ×20 multiplier begins to look like strategic infrastructure. Access would be licensed, logged and monitored. A ×100 multiplier is different in kind. No serious state would willingly share that freely with the world.
At ×100, the tool stops being a tool in the ordinary sense. It becomes concentrated national power. In commercial terms, it is the difference between entering a race on a penny-farthing and discovering that a competitor has arrived in a top-tuned Porsche. In strategic terms, it is worse, because the Porsche team can use the winnings, data and spare time to build the next vehicle.
That is why the arms-race logic appears before the hundredfold multiplier has actually arrived. States do not wait for certainty when they believe strategic advantage may be at stake. The United States can say, quite rationally, that it cannot allow China to reach the decisive multiplier first. China can say, equally rationally, that it cannot remain dependent on American models, chips or cloud infrastructure. France, or Europe more broadly, can say that it cannot become a dependency zone, renting cognition from Washington or Beijing while calling regulation sovereignty. Each argument is reasonable from inside its own perspective. Together, they produce the familiar machinery of an arms race.
This also explains why restrictions become crude. Once the feared capability is “helping people build the next AI”, it becomes very hard to draw a clean line around dangerous use. Anthropic has already shown this problem with Fable 5’s biology safeguards. In trying to avoid dangerous life-sciences assistance, the model reportedly blocked or rerouted harmless questions about ordinary biology. A similar ban on “AI development” would face the same problem. Modern AI work is made of mathematics, statistics, compilers, distributed systems, hardware optimisation, data engineering, evaluation design, security and ordinary software architecture. The boundary leaks because modern software is increasingly AI-relevant.
This is why public restrictions will not stop development. If the multiplier is valuable enough, work will continue in controlled spaces: military facilities, classified programmes, sovereign AI labs, defence contractors and heavily licensed corporate partnerships. The visible world will debate public access and responsible release. The less visible world will ask whether rivals are moving faster.
The deeper danger is that a tool no state can share cannot be governed globally in any straightforward way. Safety arguments may be sincere, but they will overlap with national advantage. Export controls may reduce real risks, but they will also shape industrial competition. Public openness may survive at lower capability levels, while the frontier moves into restricted rooms.
This is the opposite of the early promise of AI. We were told, and perhaps sincerely hoped, that artificial intelligence would make knowledge more widely available. That promise may still hold for the weaker layers. But at the frontier, the logic is changing. At ×2, AI looks like empowerment. At ×100, it looks like power.
The pattern is simple enough: every serious state wants the multiplier, but no serious state can freely share it once it becomes large enough. The race is not merely to build the best model. It is to control the tool that helps build the next one. It is not merely a race to answer questions, but to accelerate the production of answers. It is not merely a race for intelligence, but for the compounding of intelligence.
At that point, the frontier is no longer a product launch. It is a strategic boundary. And the tool no state can share becomes the tool every state must try to build.