GLM-5.2: the most powerful open-source model is Chinese and is causing concern in the West

The model also performs surprisingly well on agentic tasks and, more generally, on those most relevant to business.

On GDPval-AA v2, Artificial Analysis’s benchmark for economically relevant tasks, GLM-5.2 scores 1,524 Elo points, ahead of MiniMax-M3 on 1,418 and DeepSeek V4 Pro max on 1,328. Above all, it is broadly on a par with GPT-5.5 xhigh.

The GLM-5.2 is on a par with a Western model released around four months earlier, that is, in February 2026. In this respect, Tang Jie would be right.

Then there is the economic advantage. Z.ai puts the cost for GLM-5.2 at $1.40 per million input tokens and $4.40 per million output tokens. This is around one-tenth of the cost of US state-of-the-art models.

Claude Fable 5, on the Anthropic website, is priced at $10 per million tokens for input and $50 for output; GPT-5.5 was announced by OpenAI at $5 per million tokens for input and $30 for output.

Artificial Analysis, however, shows that GLM-5.2 is not among the most token-efficient models. In the Intelligence Index, it uses 43,000 output tokens per task, of which 37,000 are for reasoning. This is more than GLM-5.1, which used 26,000, and more than MiniMax-M3 (24,000), Kimi K2.6 (35,000) and DeepSeek V4 Pro max (37,000).

Generally speaking, this is why an increasing number of Western companies are using Chinese models for less complex tasks: the cost savings are clear.

The issue of cost is closely linked to that of access. On 12 June 2026, Anthropic announced that the US government, on national security grounds, had issued an export control directive to suspend access to Fable 5 and Mythos 5 for any foreign national, whether inside or outside the United States. The company has decided to withdraw access to the two models for all customers, and we are currently in negotiations with the US Government. The ban was triggered after the Government learnt that certain hostile state actors (Chinese and Russian) had used those models to identify some critical vulnerabilities.

This is where the ‘open weights’ approach of Chinese models makes a difference. A model that can be downloaded and managed locally reduces the risk of unilateral revocation of access by the provider or a foreign authority. However, it does not eliminate security, compliance and geopolitical issues: caution is advised.

Let’s be clear: for Western companies, using a Chinese model can mean two very different things. The first is connecting via an API to the provider’s servers, such as DeepSeek or Z.ai: in this case, texts, prompts, documents or code may leave the corporate environment and come under non-Western rules, contracts and jurisdictions. This is where the most sensitive issues arise: data protection, trade secrets, compliance, internal controls and liability in the event of misuse.

The second option is to download an open-source model and run it on the company’s own infrastructure, or with a cloud provider of the company’s choosing. In this scenario, the model may be of Chinese origin, but the data remains within the environment controlled by the company. The geopolitical risk does not disappear, as it is still necessary to assess what the model is capable of, how it has been trained and what vulnerabilities it may have.

This growing competition between the US and China remains significant for our companies. Beijing does not necessarily have to win the race outright to gain ground. It simply needs to become cost-effective, reliable and close enough to American quality standards in an increasing number of everyday applications. As, in fact, it has already managed to do. Washington retains the upper hand when it comes to the most powerful models, but risks seeing part of its application infrastructure shift towards cheaper alternatives, particularly where security and compliance do not stand in the way of adoption.

To be more specific: high-risk tasks, such as advanced cybersecurity, regulated data, sensitive scientific research, compliance and decisions with legal implications, will continue to require models with robust safeguards, audits and human oversight, as well as, in certain cases, the greatest possible processing power. Software development, refactoring, document analysis, internal automation and knowledge work, on the other hand, can be distributed across proprietary and open-source models – and thus, increasingly, across Chinese models.