How China's Z.ai Filled the Void After the U.S. Shut Down Anthropic's Fable 5 and Mythos 5

In a week that reshaped the global AI landscape, the U.S. government forced Anthropic to disable its two most powerful models, and a Chinese competitor filled the gap within 24 hours. The move has rewritten the competitive dynamics of frontier AI, ignited a stock market rally, and raised urgent questions about the export of powerful AI capabilities.

The Shutdown: June 12, 2026

On June 12, 2026, the U.S. government issued an emergency export control directive to Anthropic, ordering the company to immediately suspend all access to Claude Fable 5 and Claude Mythos 5. The directive, which cited national security authorities, barred any foreign national, including Anthropic's own foreign national employees, from accessing the models . The government's stated concern was a potential jailbreak technique that could bypass Fable 5's safety guardrails, raising fears of cybersecurity misuse .

Anthropic received the order at 5:21 p.m. ET and shut down access to both models worldwide within approximately 90 minutes . The company publicly stated it believed the order was "a misunderstanding" and was working to restore access, but the models remained offline . Other Anthropic models — including Sonnet and Haiku — were unaffected .

The Response: GLM-5.2 Arrives

On June 13, 2026, the very next day, Z.ai (formerly Zhipu AI) released GLM-5.2 to its coding-plan subscribers . The full open weights were published under the MIT License on June 16, 2026 . Z.ai's founder publicly framed the release with the message "Frontier Intelligence Belongs to Everyone" .

The timing was widely interpreted as a direct strategic response. With the top U.S. closed models pulled from global access, Z.ai offered a comparable open-weight alternative with no usage restrictions, instantly available to developers worldwide .

What GLM-5.2 Can Do

Architecture

GLM-5.2 is built on the GLM-5 Mixture-of-Experts (MoE) backbone :

• Total parameters: 744B–753B (sources vary; Together AI cites 744B, Simon Willison and others cite 753B)
• Active parameters per token: ~40B
• Context window: 1 million tokens — a 5× increase over GLM-5.1's ~200K
• Model size on disk: ~1.51TB
• Type: Text-only; a separate vision model family, GLM-5V-Turbo, exists but is not open-weight

Benchmark Performance

• ARC-AGI-2: 22.8% — the strongest open-source score, matching GPT-5.4 and GPT-5.5 under low-reasoning settings at roughly 1/7.5 the inference cost
• Coding benchmarks: Z.ai reports open-source state-of-the-art on coding and long-horizon task benchmarks, with "more stable results" in complex multi-step engineering workflows
• FrontierSWE: GLM-5.2 ranked just behind Anthropic's Claude Opus 4.8 by 1% and edged out OpenAI's GPT-5.5 by 1%
• Context retention: The 1M-token context is designed to be lossless, reducing context drift and goal forgetting in extended agentic tasks

Availability

GLM-5.2 is available through Hugging Face, Z.ai's API, Together AI, SiliconFlow, and more than 20 third-party coding tools . Users can self-host or use Z.ai's API at the same pricing as GLM-5.1 .

The Cybersecurity Dilemma

Because GLM-5.2 is released under the MIT License with no regional restrictions, it can be downloaded, modified, and deployed by anyone, anywhere — including state-aligned actors . This has raised significant cybersecurity concerns:

• No safeguards: The open weights lack the usage guardrails, content filters, and access controls that Anthropic's Fable 5 and Mythos 5 had
• Dual-use risk: The capabilities that make it strong for autonomous coding (long context, agentic execution) also make it a potential tool for generating malicious code, automating cyberattacks, or developing offensive capabilities without oversight
• Policy contrast: The U.S. government shut down Anthropic's models precisely to prevent access by foreign nationals, yet Z.ai's equivalent-capability model is freely available worldwide

Cybersecurity experts and news outlets have flagged this as a significant escalation in the open-weight risk landscape .

The Market Reaction: $128 Billion and Climbing

Z.ai listed on the Hong Kong Stock Exchange on January 8, 2026, at HK$116.20 per share with an opening market cap of approximately HK$52.83 billion (~$6.8 billion) .

After the GLM-5.2 release and the Anthropic shutdown, the stock surged dramatically:

• Market cap surpassed HK$1 trillion (~$128 billion) on June 22, 2026
• Price surge: Shares rose over 2,000% since the January IPO
• First Chinese AI company to exceed the HK$1 trillion threshold
• Dual listing plans: Z.ai is reportedly planning a secondary listing to further fuel growth

The surge was directly attributed to investor enthusiasm over GLM-5.2 filling the gap left by Anthropic's pulled models, combined with Z.ai positioning itself as a global competitor to closed U.S. frontier models .

Key Takeaways

• Timing: The U.S. shutdown created a vacuum that Z.ai filled within 24 hours, demonstrating the speed at which AI competition can shift.
• Capability: GLM-5.2 matches or exceeds several closed frontier models on key benchmarks while being freely available under the MIT license.
• Risk: The unrestricted release of such a powerful model has intensified the debate around open-weight AI and dual-use capabilities.
• Market impact: Z.ai's market cap has grown over 20-fold since its January 2026 IPO, highlighting investor belief in open-source models as a competitive force.

The episode marks a clear inflection point in the global AI race — one where a government shutdown in Washington directly accelerated the release of a competitive open-weight model in Beijing, with far-reaching implications for AI safety, policy, and market dynamics.