America Wants to Permit What China Is Starting to Restrict

TITLE: America Wants to Permit What China Is Starting to Restrict BODY: The administration that killed Anthropic's Fable and Mythos models is now quietly negotiating the opposite move for open weights. That is not a contradiction. It is a confession.

The Trump administration and the AI industry have been discussing a capability framework for US open-source models based on the current capabilities of leading Chinese open-source models. The AI industry expects Chinese Mythos-class models to be available for free download on the internet in six to twelve months. If that clock is right, the policy question stops being theoretical.

Translation: we spent two years stopping US companies from competing with China. Now we are noticing China kept competing anyway.

The Trap Is Already Closed

Under President Trump, the federal government has restricted the release of private AI models from Anthropic and OpenAI, wielding a kill-switch over models controlled by one company and based on proprietary data. Supporters of open-source models say the White House's unprecedented reach into frontier AI labs could be a blessing for China, which offers cheaper, open-source models for people and companies around the world.

The market has answered while Washington deliberates. On OpenRouter, the largest model-routing platform, Chinese open-weight models climbed from less than 2 percent of token traffic in late 2024 to 61 percent in 2026. According to Andreessen Horowitz research, 80 percent of developers around the world who use open-source tools are building with Chinese models. Those are dependency graphs. Reversing them takes years.

Zhipu's GLM-5.2, released earlier this month, can perform on par with top US labs on some cyber benchmarks, equaling Mythos' capabilities. Chinese AI models are months, not years, behind US frontier models. The Center for AI Standards and Innovation found the gap between leading US models and DeepSeek V4 Pro is about eight months.

Eight months is not a moat. It is a runway.

The effort to preserve a lead by denial may have accelerated the erosion by forcing China to build a cheaper, open, self-sufficient stack and give it away. Former Trump AI czar David Sacks wrote: "A year ago, President Trump declared that America was in a global AI race and that the way to win it was to be pro-innovation, pro-infrastructure, pro-energy, and pro-export. President Trump was exactly right; we deviate from that strategy at our peril."

The Twist Nobody Saw Coming

Just as Washington discusses permission for US open models to match Chinese capability, Beijing is discussing the opposite. Over the past month, China's Ministry of Commerce has convened Alibaba, ByteDance, and Z.ai to discuss limiting overseas access to China's top domestic AI models. The models include Alibaba's Qwen, ByteDance's Doubao, Z.ai's GLM-5.2, and DeepSeek's R1.

Officials proposed a tiered regime: basic open-source tools would require a simple filing; more capable models would face security reviews; the most sensitive frontier models would be barred from public release or restricted to domestic use only.

It is the first time Beijing has flipped the export-control logic it fought on the US chip side and pointed it at its own open-weight LLMs.

If both moves land simultaneously, the table resets. US open-weight vendors get a permission structure they currently lack. Chinese open-weight models face domestic restrictions that cost them developer distribution. If it works, the open-source share Qwen vacates gets filled inside six months by Meta's Llama and the UAE's Falcon. Speculation about Beijing's follow-through, not a guarantee. But the directional logic is sound.

On June 12, 2026, the US government suspended Fable and Mythos for foreign access, prompting Anthropic's full withdrawal. If foreign firms believe US model access can be withdrawn quickly, they will diversify. Some will choose Chinese open-weight models. Some will choose local sovereign models. Some will use multiple providers to avoid dependence on Washington.

The open-model framework discussions are the administration's attempt to offer a third option: American models with distribution guarantees baked into framework rather than subject to a 90-minute phone call.

If you are building a product on open weights, watch the threshold numbers, not the headline. The framework would set a capability ceiling based on leading Chinese models, aimed at ensuring US models maintain a strategic advantage while preventing unnecessary restrictions. A ceiling pitched low enough to exclude Llama's next release does nothing. A ceiling that matches GLM-5.2 today does not match what GLM ships in Q1. The evaluation gates tell you who is writing this framework.

What to watch: The capability threshold in any released framework and whether it references a static benchmark or a rolling Chinese model index. Whether Meta's Llama release timeline compresses to land before or after rules formalize. Whether the framework offers affirmative liability protection or merely guidance. Congressional reaction from the export-control caucus, which will read this as capitulation. And Anthropic's positioning: the company that loudest opposed open weights at frontier capability now operates under a government kill-switch it cannot control.