Chinese startup Z.ai has unveiled its latest artificial intelligence model, GLM-4.5. The company asserts that this new AI model will offer a more cost-effective solution compared to DeepSeek. The introduction of GLM-4.5 showcases how Chinese enterprises are developing advanced AI models at reduced expenses. Formerly known as Zhipu, Z.ai has confirmed that GLM-4.5 operates on "agentic" AI principles. This approach enables the model to automatically break down tasks into sub-tasks for more precise completion, a methodology distinct from existing AI models. In June, Zhipu was flagged by OpenAI in relation to Chinese AI advancements. Additionally, the US government has placed the startup on its restricted entity list, restricting American businesses from conducting transactions with it. GLM-4.5 is also open-source, allowing developers to freely download and utilize it.

What Z.ai Said About Its AI Model GLM-4.5 Being Cheaper Than DeepSeek

In a statement to CNBC, Zhang Peng, the CEO of Z.ai, disclosed that the new GLM-4.5 model operates on eight Nvidia H20 chips. These chips are Nvidia's AI training chips tailored for the Chinese market to comply with US export regulations. Although Nvidia recently received approval to resume shipments to China after a pause, the delivery timeframe remains uncertain. Zhang mentioned that Z.ai currently possesses sufficient computing resources and does not require additional chip purchases at the moment. However, he declined to reveal the training costs of the model, stating that more details would be shared later. Z.ai has priced GLM-4.5 at 11 cents per million input tokens and 28 cents per million output tokens, in contrast to DeepSeek R1's rates of 14 cents and $2.19, respectively. These tokens quantify the data volume processed by AI models. In recent weeks, various Chinese firms have introduced new open-source AI models. Tencent unveiled its HunyuanWorld-1.0 model at the World AI Conference in Shanghai, designed to aid in generating 3D scenes for game development. Following this, Alibaba launched its Qwen3-Coder model, specialized in coding tasks. Earlier this month, Moonshot, supported by Alibaba, introduced Kimi K2, which it claims outperforms OpenAI's ChatGPT and Anthropic's Claude in distinct coding tasks. According to the company's website, Kimi K2 is priced at 15 cents per million input tokens and $2.50 per million output tokens. In January, DeepSeek garnered attention from global investors with its AI model, developed despite US chip restrictions, boasting lower training and operational costs than its US counterparts. The company proclaimed that its V3 model was trained for less than $6 million, although some analysts highlighted that this figure reflects only a portion of its total hardware investment, which exceeded $500 million.