Open Weights Built China's Rise. Now Beijing May End Them.

Washington just cleared GPT-5.6 Sol for a broad public launch this Thursday — 24 hours after Beijing sat down with Alibaba, ByteDance, and Z.ai to discuss restricting overseas access to China's most advanced models. Two capitals, opposite directions, same conclusion: frontier AI is a national asset, not a product.

What broke this week:

• July 7 (Reuters): China's Ministry of Commerce held meetings with Alibaba, ByteDance, and Z.ai about limiting foreign access to frontier models — closed-source and open-weight — including models not yet released.
• July 8 (Axios): The Trump administration gave OpenAI a green light for a broad launch of GPT-5.6 Sol, Terra, and Luna this Thursday, July 9.
• June 26–July 8: GPT-5.6 spent six weeks in a government-vetted preview limited to roughly 20 trusted partners. Commerce's Center for AI Standards and Innovation ran additional safety testing; OpenAI kept technical experts in D.C.

Why Beijing is flipping:

Chinese labs built global market share by giving models away. DeepSeek's R1 shock last year proved the playbook: ship cheap, capable, downloadable weights and developers outside the U.S. switch stacks.

Reuters reports officials now worry the strategy cuts both ways:

• Mythos threat: Two sources said Chinese authorities fear Anthropic's offensive-grade Mythos could exploit software vulnerabilities against Chinese interests — echoing 360 founder Zhou Hongyi's public call for a domestic Mythos equivalent.
• Distillation panic: Alibaba banned Claude Code internally after a developer found Anthropic had slipped fingerprinting code into the tool to detect China-linked API routes. Anthropic accused Alibaba of 28.8 million distillation queries in a June letter to U.S. officials.
• National security law: Officials discussed making AI technology theft an offense under China's national security law and restricting who can fund domestic AI startups.

The tiered lock Beijing is sketching:

Reuters pointed to a May roundtable of Chinese legal experts published in a Supreme People's Court journal. The proposed framework:

• Basic open-source tools → simple filing
• Advanced models → security review
• Frontier models → barred from public release or domestic-only

Critical caveat: restrictions may apply only to future models. Qwen, Doubao, and GLM weights already shipped globally cannot be recalled.

What Washington just unlocked:

Axios reports Commerce cleared GPT-5.6 after weeks of staggered release that OpenAI said was "not its preferred way" to ship frontier models. OpenAI confirmed late Tuesday that Sol, Terra, and Luna launch publicly Thursday — initially through API and Codex, with broader ChatGPT access promised in coming weeks.

Sam Altman posted on X: "GPT-5.6 sol launches thursday! happy building."

OpenAI's own June blog post warned this guest-list process "should not become the long-term default" — but accepted it as the path to broader access while Trump's cyber executive order framework remains unfinished.

The quotes that frame the mirror:

• Carnegie Endowment fellow Scott Singer told TIME: "China will need to reckon with the reality that models that reach certain capabilities are unsafe. It is going to have the same conversations the White House has had over the last many months."
• TIME noted Z.ai's GLM 5.2 claimed to match Mythos' bug-spotting ability — and once weights are published, "they are impossible to recall or add safeguards."
• Axios: "The government and the world's most advanced AI companies are negotiating how people get access to powerful technologies case-by-case, in real time."

What stays gated:

• Mythos 5: Still limited to roughly 100 U.S. "trusted" organizations after Commerce partially lifted June export controls.
• Fable 5: Export ban lifted July 1; customer access restored within a day after new safeguards.
• Chinese frontier models: No decision made. MOFCOM and NDRC did not comment. Alibaba, ByteDance, and Z.ai did not respond to Reuters.

The market stakes:

Reuters warns any Beijing restriction could ripple globally — businesses that adopted free Chinese models to cut API costs would face price hikes. Conversely, developers locked out of Washington's guest lists have spent 2026 learning that renting American frontier models is a political risk; open weights were the hedge.

If Beijing closes that hedge on future releases, the AI stack splits into two gated ecosystems — with Europe, India, and the Global South forced to pick sides or run yesterday's weights.

Convina's view: China did not build global AI share by winning benchmarks — it built it by mailing source code while Washington debated export controls on APIs. Beijing now weighing the same gates proves the open-weight era was a tactic, not a philosophy. The irony is brutal: Commerce just cleared Sol for Thursday while China's Ministry of Commerce sketches how to keep GLM off foreign GPUs. Sovereignty in AI no longer means owning the best model. It means owning the turnstile — and both capitals just installed one within 48 hours of each other.