China's GLM 5.2 Ranks Fourth Globally as DeepSeek Adoption Rises Among US Firms After Fable 5 Export Ban

Zhipu’s GLM 5.2, released June 13, ranks as the most intelligent open-source AI model on the market and takes fourth place overall behind OpenAI’s ChatGPT 5.5 and ahead of Google’s Gemini, according to benchmarking firm Artificial Analysis. The model runs at less than a tenth the price of Anthropic’s Fable 5. Its weights are publicly available for download, as The Economist reported on June 22.

The release landed one day after the Trump administration banned non-Americans from using Fable 5 on June 12, forcing Anthropic to cut off access globally. “Our attitude is one of radical openness,” Zhipu cofounder Tang Jie said at the launch, per The Economist. He described export controls as making AI systems “subject to revocation at any moment.”

Benchmark Performance and the Gap Question

Artificial Analysis scored GLM 5.2 on dozens of benchmark tests. Fable 5 remains roughly 17% ahead on average across benchmark tasks, per The Economist. But the gap is narrower than many expected.

Havard Tveit Ihle of the Norwegian Defence Research Establishment found that Chinese models score better on public benchmarks than private ones. On public tests, Chinese models trail American ones by four to six months. On private tests, the gap nearly doubles to eight to ten months, according to The Economist. A US government study released in May found a similar gap.

GLM 5.2 shows the same pattern on two private benchmarks tested so far: seven months behind on WeirdML (unusual machine-learning reasoning tasks) and a full year behind on SimpleBench (common sense evaluation), per The Economist. But on a new Artificial Analysis exam released June 19 testing office-worker tasks, GLM 5.2 outperformed ChatGPT 5.5 despite being unable to have trained on the evaluation.

US Corporate Adoption of Chinese Models

DeepSeek saw a sharp rise in American firms paying for its services in June, according to invoicing company Ramp, as reported by The Economist. Microsoft is reportedly considering using DeepSeek’s model in its flagship Copilot chatbot.

The price gap is stark: DeepSeek charges $0.87 per million output tokens for its v4 model. Anthropic charges $50 for the same volume on Fable 5, per The Economist. But a study by Du Zheng of Georgia Tech found that DeepSeek models used 23 times more tokens than their OpenAI equivalents to achieve the same results. On a software engineering benchmark, GLM 5.2 ended up costing more than Anthropic and OpenAI systems when total token consumption was measured.

The Regulatory Asymmetry

Chinese AI models face a structural advantage in one dimension: regulatory risk. Open-source models can be downloaded and run on local hardware, outside the reach of governments or the labs themselves. The Fable 5 ban demonstrated that access to American frontier AI now rests on government approval.

The Economist noted that China’s government has not imposed similar restrictions on its own models. The absence of regulatory action may itself be evidence that Chinese authorities do not yet view their models as posing the same risks as American frontier systems.

For enterprise teams evaluating agent frameworks, the calculus is shifting. Total cost of ownership, not token pricing, determines the real expense. But supply reliability now matters as much as price: any model subject to sudden government cutoff carries a risk that open-weight alternatives do not.