I have been watching the US-China AI technology race from the inside for the past six months, and I keep coming back to the same uncomfortable question: what exactly are we locking down when we ban a model like Anthropic’s Fable 5?
Because last week a Chinese company called Zhipu AI (known as Z.ai) released GLM-5.2, an open-weight model that ties Anthropic’s Mythos on cybersecurity benchmarks, beats Claude Code on vulnerability detection, and does it at roughly one-sixth the cost. And you can download it right now.
The Numbers Are Hard to Ignore
On the CyBT-CTF benchmark, which tests how well an AI agent can investigate a security incident, GLM-5.2 scored 28 out of 59. That ties Anthropic’s Opus 4.8 exactly. It beats Claude Sonnet 4.5 by five points. And it does this while operating on hardware anyone can buy.
On Semgrep’s IDOR detection benchmark, GLM-5.2 scored a 39% F1 with no custom scaffolding, beating Claude Code at 32% and Opus 4.8 at 28%. The cost per vulnerability found: 17 cents. For a security team running scans across thousands of endpoints, that kind of per-bug economics is the difference between a technique being viable or not.
In standard coding benchmarks, it scores 81.0 on Terminal-Bench 2.1 (Opus 4.8 sits at 85.0) and 62.1 on SWE-bench Pro, edging out closed frontier models. The 1 million token context window means it can reason across an entire codebase at once.
The Distillation Question
Here is where it gets uncomfortable. Researchers at Louie.ai and Graphistry measured the correlation between GLM-5.2’s outputs and those of US frontier models. The Cohen’s Kappa between OpenAI and Anthropic models is 0.63. GLM-5.2 scores 0.80 against GPT-5.5 and 0.76 against Opus 4.8. It does not just get the right answers right. It gets the same wrong answers wrong.
That pattern strongly suggests GLM-5.2 may be a distillation or knowledge theft from US frontier models. Anthropic warned about this months ago. This is the first public measurement that shows Chinese firms have achieved frontier-level results through it.
Export Controls Just Hit Their Limit
The US government blocked Anthropic’s Fable 5 and Mythos 5 for national security reasons. The White House asked OpenAI to limit GPT-5.6. The logic was straightforward: keep the most capable models out of Chinese hands.
But GLM-5.2 is open-weight. Not open source in the full sense. The weights are published under an MIT license. You download them, run them on your own hardware, fine-tune them, and never touch a US API. Export controls on chips and models cannot stop a Chinese model that already matches the frontier and costs a fraction of the price.
The era of trying to contain AI capability through export controls ended the moment GLM-5.2 hit the download servers. Not because China won, but because the model itself is no longer the bottleneck. The data, the infrastructure, the training pipeline, those still matter enormously. But the model weights themselves are now a commodity.
What This Means
If you run a security team, GLM-5.2 running on-premises in your own environment is suddenly an attractive option. No data leaves your network. The economics work at scale. And the model quality is genuinely competitive for security work.
If you run a company that relies on US export controls as a competitive moat, you need a new strategy. The Chinese competition is not coming. It is already here, and it costs pennies.
And if you are a defender, you now face the same reality everyone in cybersecurity eventually confronts: the tools available to attackers and defenders are converging. Open-weight frontier models available to both sides mean the game is now about who uses them better, not who has access to a better model.
The US spent two years trying to build a wall around frontier AI. China just released the same capabilities as a free download. The wall was never going to hold. The only question was how fast it would fall.