China’s AI startup Z.ai (formerly Zhipu) has launched its open‑source language model GLM‑4.5, which it claims is cheaper than DeepSeek’s flagship model. Z.ai CEO Zhang Peng told CNBC that GLM‑4.5 costs just $0.11 per million input tokens and $0.28 per million output tokens, compared to DeepSeek R1’s rates of $0.14 input and $2.19 output.
How does GLM‑4.5 work?
GLM‑4.5 unifies reasoning, coding, and agentic capabilities. Developers can choose between deeper, structured reasoning or fast responses, offering flexibility for different use cases. Its architecture supports tool use and task decomposition, positioning it firmly in the emerging agentic AI ecosystem.
Z.ai’s GLM‑4.5 will come in two variants, GLM‑4.5 and GLM‑4.5‑Air, designed for “agentic” applications. These models can autonomously break down complex tasks into subtasks when in “thinking mode” and deliver quick responses via a simpler “non‑thinking” pathway. Importantly, the model runs on just eight Nvidia H20 GPUs and is fully open‑source, allowing developers to download and deploy it freely.
Impact on DeepSeek and competitors
DeepSeek shook the global AI community earlier this year with R1 and V3 models trained at dramatically lower cost using sparse MoE and efficient inference. GLM‑4.5 raises the bar yet again, low-balling DeepSeek’s pricing while offering comparable capabilities. This intensifies competition among Chinese firms, even as Alibaba’s Qwen and Baidu’s Ernie models target reasoning tasks at lower cost.
Industry experts highlight GLM‑4.5’s cost‑efficiency and agentic framework as meaningful advances. Observers emphasize that this model reflects China’s broader push to deliver powerful open‑source AI at minimal cost, looking to disrupt Western competitors’ pricing models.
Z.ai, rooted in the former Zhipu AI, is backed by investors such as Alibaba, Tencent, and local government funds. The startup has raised over $1.5 billion and is reportedly preparing for an IPO in China.