Beijing — Chinese AI startup Z.ai has announced the release of its latest language model, GLM-4.5, claiming it to be more cost-effective than its domestic rival DeepSeek. The company says the model is open source and free to download, while offering lower operational costs.
Previously known as Zhipu, the company revealed that GLM-4.5 requires just eight Nvidia H20 chips to function. These chips, designed specifically for the Chinese market under U.S. export rules, offer a key advantage in minimizing hardware usage. Z.ai CEO Zhang Peng said the company currently has sufficient computing power and will share the training costs for the model at a later stage.
The model is based on “agentic” AI, a system where the model independently breaks down tasks into smaller steps to improve accuracy. This design differs from traditional large language models, and aims to deliver more efficient and accurate results.
Z.ai also announced aggressive pricing: just 11 cents per million input tokens and 28 cents for output tokens, significantly lower than DeepSeek’s rates of 14 cents and $2.19 respectively. Tokens are the units used to measure data processed by AI models.
The release of GLM-4.5 comes as competition intensifies among Chinese tech firms to build powerful AI tools at lower costs. Earlier this month, Moonshot AI, backed by Alibaba, launched Kimi K2, another challenger to ChatGPT, which charges 15 cents for input and $2.50 for output per million tokens.
Z.ai’s announcement follows DeepSeek’s own global attention earlier this year, when its V3 model surprised markets with low training costs and performance comparable to top U.S. models. While DeepSeek’s V3 was reportedly trained for under $6 million, experts noted the company had previously invested over $500 million in hardware.
The race to develop faster, cheaper AI in China continues amid ongoing chip restrictions and strategic competition with U.S. firms. With GLM-4.5 now in the open source space, developers and researchers worldwide will be watching closely to see how it performs in real-world applications.
