China’s Z.ai unveils GLM-5.3, claims chart-leading scores

Z.ai, a Chinese AI company, has unveiled its latest open-weights model GLM-5.3, which it claims outperforms both Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol on the CyberGym benchmark. The Hong Kong-listed firm says the new model rivals the capabilities of the most recent launches from US AI companies, marking a fierce competition in the race for AI supremacy.

GLM-5.3's performance is a significant improvement over its predecessor GLM-5.2, which was launched in June. The company touts a 50% boost in its in-house code benchmarks with the new model. Z.ai plans to release the weights of GLM-5.3 in two weeks after completing safety evaluations and enhancements. The model demonstrates superior agentic coding abilities, scoring 84.5% on CyberGym, outperforming both Anthropic's and OpenAI's models.

Despite the impressive claims, Z.ai's shares dropped nearly 4% in value by market close today, with intelligence analyst Robert Lea warning that the firm remains on an unsustainable commercial footing due to the rising costs of agentic AI. Chinese AI models continue to match and sometimes surpass American counterparts in performance at lower prices, even with restrictions on chip exports into the country.

In July, Z.ai completed the construction of a new data center housing at least 10,000 Chinese-made chips to support its GLM model line. Other notable Chinese AI offerings such as Moonshot's Kimi model, DeepSeek's V4 models, and Alibaba's Qwen series have also performed well in benchmarks and tests. Z.ai went public in January, raising nearly $7bn in funding, and its shares surged 2,000% in June following the launch of GLM-5.2, pushing its valuation to $128bn before it settled around $75bn.