Anthropic made its most advanced AI model, Fable, generally available earlier in June, sparking a cautionary tale for American AI companies.
Three days after launch, the company suspended access for its paid users over concerns that hackers had circumvented the model’s built-in guardrails. Last Friday, a select group of American companies regained access after receiving U.S. government clearance. Then, on June 30, Anthropic said it would reopen Fable — a trimmed version of its flagship language model, Mythos — to a limited group of users. Pro and Max subscribers will have to wait another week, and even then, they’ll get a version with reduced capabilities.
Its rival, OpenAI, avoided a similar disruption with GPT-5.6 by restricting access from the outset, releasing the model only to a government-vetted list of users.
For most of the past two years, Washington’s approach to artificial intelligence has been simple: let the labs run. That changed on June 12, when the Commerce Department classified Anthropic’s newest models as dual-use cyber tools and ordered every foreign national, including Anthropic’s own foreign-born employees in San Francisco, to lose access. With no quick way to verify citizenship, Anthropic effectively shut the models off for everyone.
Twist in the tale
The administration has said the restrictions followed the discovery of a guardrail bypass identified by Amazon security researchers. But that explanation landed alongside reports that a China-linked operation had been quietly extracting capabilities from Anthropic’s systems.
That wasn’t the company’s only headache. Anthropic also told the U.S. Senate Banking Committee that accounts linked to Alibaba’s Qwen generated nearly 29 million exchanges through roughly 25,000 fake accounts over six weeks, in what it described as the largest known attempt to distill its models into a cheaper rival.
At the same time, the Pentagon had separately blacklisted Anthropic as a supply-chain risk after the company declined to make its tools available for mass surveillance and autonomous weapons. The result was a strange contradiction: a company judged too risky for parts of its own government to use, while simultaneously being deemed too dangerous to leave broadly accessible.
Rivals proliferate
Less than two years ago, DeepSeek demonstrated that it could approach American AI frontier performance on a fraction of the budget. This time, a day after Washington pulled the plug on Anthropic’s flagship models, Beijing-based Z.ai released an open-weight model that scored within a point of Anthropic’s best on a closely watched coding benchmark while reportedly costing around a fifth as much to run—and doing so on Chinese-made chips.
Z.ai’s co-founder didn’t bother hiding the comparison, calling the American suspension “deeply regrettable” and arguing that frontier intelligence shouldn’t disappear because of a government letter.
China is not the only contender this time. In Japan, Sakana AI launched its own frontier model, marketing it as delivering cutting-edge capability without the risk of export controls. A Chinese cybersecurity firm released a competing cyber-focused model the very same week.
Once a frontier model has been trained and its weights released — as several of China’s strongest models now have been under permissive open licenses — no government order can pull them back off the internet. Export controls work because hardware is scarce. But software is not.
Going back in time
In 1945, the United States held a monopoly on the only weapon that mattered and worked hard to preserve it. Washington proposed placing atomic energy under international control through the Baruch Plan, while the 1946 Atomic Energy Act erected such sweeping secrecy that even Britain and Canada — wartime partners that had helped build the bomb — were shut out. American officials convinced themselves that the Soviet Union was a decade away.
But the Soviets detonated their first nuclear device in August 1949, just four years after Hiroshima, years ahead of American expectations. They got there through a combination of capable domestic science and successful espionage. Secrecy had not stopped the rivalry. It had mostly alienated America’s allies as Britain built its own independent nuclear deterrent rather than depend on technology that could be withdrawn at Washington’s discretion.
A question of strategic risk
The immediate consequence of the present export restrictions has not stopped Chinese firms from competing. Instead, companies and governments elsewhere are beginning to ask whether dependence on American frontier models is itself becoming a strategic risk.
Japan’s Sakana AI has already made export-control resilience part of its sales pitch. European policymakers are reportedly weighing new AI sovereignty investments as businesses cannot build using technology that might disappear overnight.
So the lesson from 1949 is that restrictions may buy time, but they can stop competitors. The longer-lasting cost often falls not on the rival but on the country doing the restricting.
American AI companies are heading to public markets on the promise that the world’s most capable models will continue to come from the U.S. That promise becomes harder to sustain once “best” comes with a ‘conditions apply’ label.
Between Mythos and Fable, America’s frontier AI strategy increasingly seems caught between openness and control. The question is whether that balance slows competitors or simply encourages the rest of the world to build something it cannot lose access to.
