Z.ai, formerly known as Zhipu, has unveiled its new GLM‑4.5 model, aiming to disrupt the AI market by undercutting DeepSeek R1 in pricing while maintaining high performance.
At the World AI Conference in Shanghai on July 26, 2025, CEO Zhang Peng announced that GLM‑4.5 will be significantly more affordable than DeepSeek R1. The open-source agentic model uses only eight Nvidia H20 chips to run, compared with DeepSeek’s much larger infrastructure
Key Features of GLM‑4.5
- Cost advantage: Input token price estimated at $0.11 per million tokens and output token price at $0.28, vs DeepSeek’s $0.14–$0.55 input and $2.19 output
- Model size & agentic design: Approximately half the size of DeepSeek’s model, GLM‑4.5 is optimized for intelligent task decomposition and efficiency.
- Open-source licensing: Released under an MIT license, free for download and commercial use.
DeepSeek R1: The Benchmark Under Threat
DeepSeek’s R1 model gained attention in early 2025 for delivering competitive math, coding, and reasoning capabilities while keeping costs extremely low.
- API pricing ranges from $0.14 to $0.55 input per million tokens, and $2.19 output per million tokens
- R1 was trained using roughly 2,000 Nvidia H800 GPUs and cost around $5.6 million—vastly more efficient than competing models trained on H100s or H200s
- Despite its low cost, R1 performed on par or better than OpenAI’s o1 in key benchmarks like MATH-500 and SWE‑bench
Why GLM‑4.5 Matters
- Lower operational cost: GLM‑4.5’s pricing could reduce AI inference costs further, especially for developers and enterprises deploying at scale.
- Agentic capabilities: The model automatically breaks down tasks into sub-tasks for better accuracy and efficiency—a capability emerging as a key differentiator in new-generation models
- Competition in China: Z.ai’s launch follows other recent Chinese releases like Moonshot’s Kimi K2 and Tencent’s HunyuanWorld‑1.0, illustrating rising competition in open-source AI from China CNBC.
Industry Reactions
Analysts suggest GLM‑4.5 could trigger a price war:
- Token economics: With input pricing ~20–95% lower than DeepSeek and output pricing over 7× cheaper, GLM‑4.5 directly challenges R1’s value proposition.
- Open-source momentum: Following DeepSeek’s example, Z.ai reinforces the trend of MIT-licensed, developer-friendly models reshaping AI accessibility.
- Strategic positioning: Backed by major investors including Alibaba and government funds, Z.ai is positioning itself as one of China’s “AI tigers” competing globally.
Summary Table
Feature | Z.ai GLM‑4.5 | DeepSeek R1 |
---|---|---|
API Input Cost | ~$0.11 per million tokens | $0.14–$0.55 |
API Output Cost | ~$0.28 per million tokens | $2.19 |
Model Size | ~half of DeepSeek’s | Full R1 scale (~671 B params) |
Compute Requirement | 8 Nvidia H20 chips | ~2,000 Nvidia H800s |
License | MIT open-source | MIT open-source |