Open-source AI models are rapidly shifting the economics of artificial intelligence, with China’s Z.ai introducing GLM-4.5—a model that redefines what’s possible in terms of both cost and performance. Released in July 2025, GLM-4.5 is designed to execute complex reasoning, coding, and multimodal tasks while requiring significantly fewer resources than previous models, marking a pivotal moment in the AI race between China and the US.

GLM-4.5’s standout feature is its aggressive pricing: input processing costs just $0.11 per million tokens and outputs just $0.28 per million tokens. This is not only lower than China’s previous disruptor, DeepSeek, but also drastically undercuts leading Western models such as Anthropic’s Claude and OpenAI’s GPT series. For context, Claude 3 Opus charges up to $15 per million output tokens—GLM-4.5’s rate is more than 50 times lower, making advanced AI accessible to a wider range of users and businesses.

Technical advances drive this cost reduction. GLM-4.5 uses a model architecture with 355 billion parameters but only activates 32 billion at any one time, which trims hardware requirements without sacrificing performance. The model operates efficiently on just eight Nvidia H20 GPUs—a chip specifically tailored for the Chinese market to comply with US export restrictions. This compact design slashes both infrastructure costs and energy consumption, allowing smaller organizations to deploy powerful AI without the need for massive data centers.

Training for GLM-4.5 followed a staged approach. Z.ai initially built the model using a dataset of 15 trillion tokens, then refined its reasoning abilities with an additional 7 trillion tokens. This focus on “agentic” AI—where the model automatically breaks down complex tasks into sub-tasks—yields improved accuracy and reliability, particularly for applications demanding advanced reasoning or multi-step problem-solving.

GLM-4.5’s Open-Source Accessibility

GLM-4.5 is available for free download via platforms like GitHub and Hugging Face, and can also be accessed through Z.ai’s browser-based interface. This open-source approach accelerates adoption, giving developers, startups, and researchers direct access to a cutting-edge model without licensing fees or restrictive terms. The release includes not only the flagship GLM-4.5, but also GLM-4.5-Air (a smaller, more resource-efficient variant) and GLM-4.5-Flash (optimized for coding and agent tasks), offering flexibility for a range of use cases and hardware environments.

Early demand for GLM-4.5 was so high that Z.ai’s public demo reached capacity within hours of launch, highlighting significant interest from the global AI community. The model’s performance on industry benchmarks places it just behind leaders like OpenAI’s o3 and xAI’s Grok 4, while outperforming many established alternatives on cost-sensitive tasks.

Comparing GLM-4.5 to DeepSeek and US AI Models

DeepSeek previously set a new standard for cost efficiency, claiming to train its R1 model for under $6 million and operate it on a fraction of the hardware used by OpenAI. GLM-4.5 builds on this by halving model size and further reducing chip requirements. Input and output costs for GLM-4.5 are lower than DeepSeek’s R1, and dramatically less than those of US-based models, which often require costly, high-end GPUs and sprawling infrastructure.

While US firms like OpenAI, Anthropic, Meta, and Google pursue ever-larger, more expensive models, Chinese companies are focusing on streamlining architectures and optimizing for cost. This divergence is reshaping global AI competition, as open-source, low-cost models become "good enough" for many business and research needs, eroding the market advantage of expensive proprietary systems.

Considerations and Cautions for Adoption

Despite its technical strengths and cost advantages, GLM-4.5’s adoption in the West may be limited by regulatory and privacy concerns. Z.ai, formerly known as Zhipu, is on the US Commerce Department’s Entity List, restricting American companies from collaborating or transacting with it. Additionally, questions around data privacy and potential censorship have been raised, especially for users with sensitive or regulated data. Organizations considering GLM-4.5 should consult compliance teams and assess risks before integration.

For marketers, startups, and small businesses, GLM-4.5’s open-source license and low operating costs make it an appealing alternative to established platforms. Its ability to process large volumes of data and perform advanced reasoning can level the playing field for smaller players, provided they remain vigilant about compliance and data governance.

Alternative Open-Source AI Models in China

GLM-4.5 is part of a broader surge in open-source AI development in China. Alibaba’s Qwen3-Coder and Moonshot’s Kimi K2 have also entered the field, each offering competitive performance and pricing. Tencent’s HunyuanWorld-1.0 targets 3D scene generation, further diversifying the landscape. The sheer volume—over 1,500 large-language models released in China as of July 2025—demonstrates a strategic focus on rapid, cost-effective AI innovation.

GLM-4.5’s debut signals a new phase in AI competition, where cost, efficiency, and open access are becoming as important as raw performance. As open-source models continue to improve, expect ongoing shifts in how organizations choose and deploy AI solutions worldwide.