Chinese startup Z.ai announced on Monday that its new open source AI model GLM-4.5 will be less expensive to use than DeepSeek. In addition, GLM-4.5 is reportedly about half the size of DeepSeek and can be freely downloaded and used by developers.
The company, formerly known as Zhipu, said the new GLM-4.5 is its first foundation model built on agentic AI, which is a technology designed to autonomously break down complex tasks into manageable subtasks to enhance execution accuracy.
Z.ai CEO Zhang Peng stated in the press release: “We are setting a new benchmark with GLM-4.5, demonstrating that cutting-edge performance can be open, efficient, and affordable. The first-principles approach to measuring AGI is to integrate more general intelligent capabilities without losing existing ones. GLM-4.5 is our first complete realization of this concept.”
GLM-4.5 is already a standout in China’s AI space
GLM-4.5 is Z.ai’s debut into open-source development with a Mixture of Experts (MoE) model design. It comes in two configurations: the 355-billion-parameter flagship GLM-4.5 and more compact 106 billion-parameter GLM-4.5-Air. In internal evaluations across 12 representative benchmarks, the model reportedly ranked third globally and led among Chinese and open-source competitors.
Z.ai emphasized the model’s operational efficiency, stating it requires only eight NVIDIA H20 chips — i.e., hardware specifically designed for China under US export restrictions — according to Zhang’s interview with CNBC.
Earlier this month, NVIDIA confirmed it would resume chip shipments to China following a three-month pause, contingent upon license approvals. Zhang told CNBC that Z.ai doesn’t need to purchase any more chips, as it already has enough computing power.
DeepSeek’s training costs and rates
DeepSeek made a significant impact in the AI scene in January with its ability to create an affordable model that rivaled OpenAI’s ChatGPT, despite limitations due to US chip restrictions. The company claimed that its training costs were less than $6 million for the V3 model, although some industry analysts speculate the figure was based on the company’s hardware spend of more than $500 million over time.
Zhang declined to provide any information on the amount spent on AI model training and stated that more details regarding the new GLM-4.5 model would be released later.
Z.ai plans to charge $0.11 per million input tokens and $0.28 per million output tokens for GLM-4.5 usage, undercutting DeepSeek’s respective rates of $0.14 and $2.19.
More AI development in China leads to OpenAI’s warning
Founded in 2019, Z.ai has focused on advancing AI technologies and has raised more than $1.5 billion from investors. As a result, the company was added to the U.S. Entity List, which restricts American firms from doing business with listed entities without special government approval. In June, OpenAI issued a warning about the pace of Chinese AI development, naming Z.ai — then operating as Zhipu — as one of the companies to watch.
Z.ai’s new model is just one of the many open-source AI models being produced by Chinese developers.
- Earlier this month, Moonshot, an Alibaba-backed company, introduced its Kimi K2 model, which it claimed performs better than ChatGPT in specific coding benchmarks.
- Tencent introduced its HunyuanWorld-1.0 model at the World AI Conference in Shanghai. The model is designed to generate 3D scenes for gaming environments.
- Last week, Alibaba announced its new Qwen3-Coder model for performing programming tasks.