Last month, the U.S. Commerce Department ordered Anthropic to cut off access to the company’s two most advanced AI models. Within a day, a Chinese lab answered by releasing a competing model of its own for free. That frames the recent hair-raising AI standoff between the two countries.

Then, yesterday, Washington blinked, withdrawing the controls. Anthropic said it would begin restoring access today.

Open source and closed models have become competing strategic narratives in the global competition for AI dominance. It's hair-raising because the stakes are so high: whichever side wins will dominate AI deployment.

Anthropic and OpenAI keep their most powerful models proprietary and gated behind APIs, arguing that closed access makes them easier to monitor and keep aligned with Western norms. That is a commercial and safety choice. The export ban was something else: the government reaching for control over who may touch a technology, on national-security grounds. China is pushing the opposite: open weights that any developer can download and run locally.

If you're feeling some cognitive dissonance, you're not alone. The champion of open markets reached for state control over who may use a private company's product, while the government most obsessed with centralized control is handing its most powerful technology to the world.

While the closed models from the U.S. have maintained a performance edge over China's open models, that gap is closing fast. What is more, China's open models are cheaper to run and easier to customize. The world is watching like spectators at a tennis match, swiveling back and forth as governments and enterprises decide which side to rely on.

Washington is betting that gatekeeping can contain the risks even if it slows global adoption. China is betting that open source will turn its models into de facto standards worldwide, taking market share from America's closed models and making it harder for Washington to wall off advanced AI without pushing users toward Chinese alternatives.

The tug of war is also hair-raising because China's approach puts the world's most powerful AI in the hands of bad actors, who could – and likely will – use the models to wreak havoc.

We need an international framework to set some rules.

On June 12, the Commerce Department ordered Anthropic to suspend access to its new Fable 5 and Mythos 5 models for every foreign national, citing national security authorities. The order covered not only people abroad but non-citizens inside the United States, including Anthropic's own employees. Because the company had no way to verify nationality inside a live product, it shut down both models worldwide within hours.

Officials treated the move as preemptive containment. But the practical result was that the country had cut its most capable commercial models off from most of the world.

About a day after the shutdown, the Chinese firm Zhipu, which operates as Z.ai, released GLM-5.2 under an MIT open-source license, with open weights and a standalone API to follow within the week. Founder Jie Tang declared that advanced AI should be open to anyone, not belong to a small group, or be pulled back at will.

GLM-5.2 did not come out of nowhere. Chinese open models had been gaining ground for months: Moonshot's Kimi, Alibaba's Qwen line, DeepSeek's cut-rate systems. Independent analysts now rate GLM-5.2 the most capable open-weight model available.

The logic behind the export order was straightforward. If Fable 5 or Mythos 5 could be jailbroken and pointed at offensive cyber work, keeping them away from adversaries was a national-security imperative. The trouble is that model weights and techniques don't stay contained for long in an industry moving this fast, and the ban fell on allies and ordinary customers as heavily as on any adversary.

By shutting the models off from most of the world, the ban handed Chinese labs a ready-made line: U.S. AI is locked up for insiders, Chinese AI is open to everyone. That pitch appeals to developers in the Global South and to cost-conscious companies in the U.S. and Europe, who see cheap, self-hostable Chinese models as a way around both export limits and steep cloud bills.

Eighteen days after the shutdown, Washington reversed itself: Commerce Secretary Howard Lutnick said a license would no longer be required to export Anthropic’s models. In return, Anthropic agreed to detect and address security risks itself and to report malicious activity to the government. No one mentioned China.

It was as much an admission as a concession: hard limits on model access are difficult to hold when comparable systems are being open-sourced from Beijing. GLM-5.2 did not set U.S. policy, but it raised the cost of keeping Anthropic's models offline.

None of this is settled. More capable models will bring more pressure to control them, and more reason for rival governments to respond to a lockdown by opening their own systems. As models grow more powerful, that dynamic is likely to sharpen. Managing increasingly dangerous AI will require international cooperation on shared standards and monitoring, not just national gatekeeping. It's time for both sides to talk.