China's 'Next DeepSeek Moment'

GLM 5.2 was developed by the Chinese company Zhipu AI (Z.ai) and is attracting attention thanks to programming capabilities nearly equivalent to today's most powerful American models.

According to the FrontierSWE rating, which measures how well an AI can complete large-scale engineering projects lasting dozens of hours, GLM-5.2 scored 74.4%, surpassing OpenAI's GPT-5.5 and Anthropic's Opus 4.7. China's new open source model is less than a percentage point behind Opus 4.8.

In two other ratings, PostTrainBench and SWE-Marathon, which both measure the ability to perform complex and multi-step tasks, GLM-5.2 is in the top 3 globally along with the best US models.

"GLM-5.2 marks a new 'DeepSeek moment' for Chinese AI," Kyle Chan, a researcher at the Brookings Institution, said on SCMP. "Z.ai's new model shows that China is only a few months behind the US in AI performance, despite having much less access to computing power."

According to Chan, this open source model comes at a "perfect period" when many people are concerned about skyrocketing costs for AI. The cost of using the GLM-5.2 is half that of models from the American company, while the capacity is at the top.

Programming is one of the fastest growing commercial applications of the major language models. Anthropic, a leader in programmable AI, last month projected annual revenue of about $47 billion, up from $1 billion in early 2025, according to recent funding documents.

According to Nathan Lambert, a researcher at the Allen Institute for Artificial Intelligence Research, GLM-5.2 is one of the first open weight models that can be considered as an alternative. This type of model allows the user to adjust the set of trained weights.

However, according to CNBC, the GLM-5.2 is still behind the US model in some reviews. DeepSWE, a ranking that evaluates software engineering capabilities, places this model in 5th place, behind models from Anthropic and OpenAI.

Users also noted some operational inefficiencies. Sridhar Ramaswamy, CEO of American data software provider Snowflake, said that GLM-5.2 achieves overall performance equivalent to Claude Opus 4.7, but exhibits some limitations such as stopping tasks too early and over-analyzing inaccurate details.

Tang Jie, founder of Zhipu AI, shared on social media that the model could improve if more resources were devoted to reinforcement learning.

Nam Nguyen