Summary
Imagine you have been using a very powerful calculator for your business — one that costs $1,000 a month and is made by a company in California.
Then, one morning, you wake up to find that the California company has been told by the government that it cannot sell its calculator to people outside the United States anymore.
The same morning, a company in Beijing announces that it has made a calculator that is almost as powerful — and it is giving it away for free, to anyone in the world, with no restrictions.
That is, roughly speaking, what happened on June 13, 2026, when a Chinese company called Z.ai released a powerful artificial intelligence model called GLM-5.2, one day after the United States government banned access to Anthropic’s most advanced AI models for people outside America.
GLM-5.2 is what experts call an open-weight model.
Think of it like a recipe. Most big American AI companies keep their recipes secret — you can use the food they make, but you cannot see how it was cooked, copy it, or change the ingredients.
GLM-5.2 is different. Z.ai has published the entire recipe under an MIT license, which is a legal agreement that says anyone in the world can download it, run it on their own computers, change it, improve it, and use it commercially — all without asking Z.ai’s permission and without paying a cent. The model’s weights — its core mathematical instructions — are sitting on a website called Hugging Face, available to download right now.
The model is very large: it has 744 billion mathematical parameters, which is roughly the scale of the largest American AI systems. It uses a technique called mixture-of-experts, which means that while it has hundreds of billions of parameters in total, only about 40 billion of them do the work on any given task — similar to how a large hospital has hundreds of specialist doctors, but you only see the relevant ones during your visit. This keeps inference costs low even though the model is vast.
What makes GLM-5.2 genuinely remarkable is what it can do with code and complex, multi-step tasks. Independent testing organisations found that it scores higher than every other openly available model on the most important coding benchmarks.
On one key test called SWE-bench Pro, which challenges AI systems to fix real software problems, GLM-5.2 scored 62.1 — better than OpenAI’s GPT-5.5, which scored 58.6. It still trails Anthropic’s Claude Opus 4.8 on the hardest long-horizon tests, like a benchmark called SWE-Marathon, where it scored 13.0 against Opus 4.8’s 26.0.
But the gap is narrowing fast, and for most practical business and engineering uses, the performance difference is small.
The price difference, however, is enormous. Z.ai charges approximately $1.40 per million words of input and $4.40 per million words of output.
Anthropic’s equivalent service costs $10 per million words of input and $50 per million words of output. OpenAI’s GPT-5.5 costs $5 per million words of input and $30 for output.
In plain terms: you can do the same coding or research task with GLM-5.2 for roughly one-tenth of what it costs with the leading American models. A company that was spending $10,000 a month on AI can now potentially spend $1,000 to $2,000. Coinbase, a large American crypto company, publicly announced it had switched to using GLM-5.2 and another Chinese model, cutting its AI spending by nearly half.
There is another detail that has shaken the technology and foreign policy worlds: GLM-5.2 was trained entirely on Chinese-made computer chips, with no American hardware involved.
The United States government has spent years trying to prevent China from buying Nvidia’s most advanced chips — the chips that power AI training — because it believed that without those chips, China could not build truly powerful AI.
Z.ai built GLM-5.2 on a cluster of approximately one hundred thousand Huawei Ascend 910B chips, which are designed by a Chinese company and manufactured in China.
The training was slower and less efficient than it would have been on Nvidia hardware, but the result is still a frontier-class model.
This has been described by analysts as the moment that America’s semiconductor export control strategy received its clearest practical rebuttal.
Dr. Antonio Bhardwaj, a globally recognised expert in Human-Centered AI for Geopolitical Strategy and biohazard risk, put it this way: “The release of GLM-5.2 demonstrates that the long-standing assumption underlying Western technology policy — that hardware restrictions could function as a reliable ceiling on Chinese AI capability — has now been empirically disproved. The model exists, the weights are distributed globally, and the containment architecture built around supply chain restriction has reached the limits of its effectiveness.”
The timing of the release carries its own message.
The American government banned Anthropic’s Fable 5 and Mythos 5 models for all foreign users on June 12, 2026, citing concerns that hostile actors had used the models to find dangerous software vulnerabilities.
The very next day, Z.ai announced GLM-5.2 — a model that, according to two independent cybersecurity firms, can itself find software vulnerabilities at a level comparable to those restricted American models.
Unlike the American models, though, there is no government that can ban GLM-5.2, because it is already downloaded onto computers around the world. Security researchers have confirmed that within days of the release, hackers were sharing methods to remove the model’s safety restrictions on Russian-language online forums.
Dr. Bhardwaj has also raised a warning that extends beyond cybersecurity: “The autonomous reasoning capabilities in models like GLM-5.2 are not limited to software engineering. A system capable of sustained, multi-step agentic work can, in principle, assist in the design of dangerous biological agents or navigate scientific literature for dual-use research pathways. Treating open-weight frontier models solely as a coding or commercial challenge misses the biohazard dimension entirely.”
For people outside the United States who had built businesses or research programmes on Anthropic or OpenAI’s models, the week of June 13, 2026 was a turning point.
The Fable 5 ban showed that relying on American AI could mean losing access overnight if the US government decided it was a national security risk. GLM-5.2 offered an alternative that nobody could take away — because the recipe is already in your hands.
Z.ai went public on the Hong Kong Stock Exchange in January 2026, and after the GLM-5.2 release its stock more than doubled in a single week, reaching a market value of over HK$650 billion.
Major banks raised their price targets. The company is preparing a second stock market listing in Shanghai.
None of this means GLM-5.2 has beaten the best American models on every task. On the hardest reasoning problems, significant gaps remain.
The model also lacks the ability to process images — a meaningful limitation compared to Anthropic’s and OpenAI’s latest offerings. And anyone who routes their data through Z.ai’s cloud service rather than running the model on their own hardware should be aware that Chinese law can require Chinese companies to share data with the government on demand.
But the overall direction of the story is clear. Two years ago, China was building AI models that were noticeably behind the American frontier.
Today, on many practical tasks — coding, long-horizon planning, software engineering — a freely available Chinese model sits within a few percentage points of the best American systems, at one-tenth the cost, on Chinese-made chips, with a license that no government can revoke.
Dr. Bhardwaj’s conclusion is sombre and direct: “The global AI competition has entered a phase where the key variable is no longer who has the most powerful model in an absolute sense, but whose model is embedded most deeply in the world’s infrastructure, workflows, and institutions. GLM-5.2’s open-weight strategy is a deliberate and sophisticated bid to win that embedding contest. The world’s governments and international institutions need to respond with governance frameworks commensurate with the challenge — and they need to do so quickly.”
The frontier has shifted. The question now is whether the rules that govern it will shift fast enough to keep pace.
