U.S. Restrictions on Anthropic’s Frontier Models Accelerate Chinese Startup Z.ai’s GLM-5.2 Entry

On June 12 2026 the U.S. government ordered Anthropic to suspend its Fable 5 model, citing security concerns that could expose the Mythos model. The order limited access for foreign nationals, prompting a temporary shutdown of both models before limited Fable 5 access was restored under stricter nationality controls. Startup Legion sued over the restriction.

Anthropic's distillation problem reveals that export controls alone cannot hold the line in the US-China AI raceNew Chinese AI GLM-5.2 Beats Every ChatGPT Model, Trails Only Anthropic's Claude FableNew York Times

Anthropic reported that DeepSeek, Moonshot AI and MiniMax used about 24,000 fake accounts and proxy services to bypass Claude restrictions, generating over 16 million exchanges: DeepSeek >150 k, Moonshot AI 3.4 million, MiniMax >13 million. Anthropic called the effort a distillation that evades chip-level export controls.

Anthropic's distillation problem reveals that export controls alone cannot hold the line in the US-China AI race

• Feb 2026: Anthropic discloses Chinese distillation.

Anthropic's distillation problem reveals that export controls alone cannot hold the line in the US-China AI race

• Jun 12 2026: U.S. suspends Fable 5.

Anthropic's distillation problem reveals that export controls alone cannot hold the line in the US-China AI race

• Mid-Jun 2026: Z.ai launches GLM-5.2, a top-three model.

New York Times

Benchmarks rank GLM-5.2 third, behind Anthropic’s Claude Mythos and Claude Fable and ahead of OpenAI’s ChatGPT. Pricing is about $1.40 per million input tokens versus roughly $15 for Claude Opus 4.8. The open-weight license lets developers download, modify and run the model on private hardware, enabling a $10 000-per-month closed-source team to achieve similar output for far less.

New Chinese AI GLM-5.2 Beats Every ChatGPT Model, Trails Only Anthropic's Claude Fable

GLM-5.2’s low price and open access have drawn Silicon-Valley developers seeking cheaper AI. U.S. firms warn of ties to the Chinese government and data-security risks when using foreign servers. Critics note that restricting American models may unintentionally increase market share for cheaper Chinese alternatives.

New York Times

Anthropic linked the fraudulent activity to the three Chinese firms through technical fingerprints. Senator Mark Warner cited a test where Mythos found vulnerabilities in classified systems, calling it a controlled red-team exercise. Z.ai founder Tang Jie rejected claims that parity with U.S. models is a year away, saying equivalence will arrive sooner. DeepSeek founder Liang Wenfeng said chip access, not funding, is the main obstacle for Chinese labs.

Anthropic's distillation problem reveals that export controls alone cannot hold the line in the US-China AI race

U.S. firms accuse Chinese rivals of using tens of thousands of fake accounts and nearly 29 million queries to harvest model outputs. Anthropic’s figure of “more than 16 million” exchanges differs from the “nearly 29 million” cited elsewhere, highlighting a data discrepancy. Developers also remain hesitant to route sensitive data to Chinese servers, noting that the hardest reasoning tests still favor American labs.

New Chinese AI GLM-5.2 Beats Every ChatGPT Model, Trails Only Anthropic's Claude Fable

> "With Fable restricted, the gap between the U.S. and China is very slim." — Rehaan Ahmad, co-founder, alphaXiv

New York Times

Legal challenges such as Legion’s lawsuit and U.S. export-control discussions suggest continued regulatory pressure. Meanwhile, Chinese developers are adapting models to run on domestic chips, positioning GLM-5.2 and similar systems as cost-effective alternatives for AI users.

New York Times