Beijing-based Z.ai has released GLM-5.2, a 744-billion-parameter mixture-of-experts model that ranks among the top five AI systems globally while carrying a price tag roughly one-sixth that of leading US frontier models.
What the model does
GLM-5.2 is designed for long-horizon coding and agentic tasks. It achieved second place on Code Arena (1,595 points), first place on Design Arena, and topped the open-model tier on Agent Arena, according to the company. On the SWE-bench Pro coding evaluation it scored 62.1, compared with 69.2 for Anthropic’s Claude Opus 4.8. The Artificial Analysis Intelligence Index v4.1 rates it at 51, ahead of Google’s Gemini 3.1 Pro Preview at 46.
The model offers a one-million-token context window, quadruple that of its predecessor, and operates in two reasoning modes for varied task demands.
Open weights, no regional locks
Z.ai released the model weights under the MIT licence with no usage restrictions and no regional limits. The company frames this as technical access without borders, a pointed contrast to US export controls that temporarily blocked Anthropic’s Fable 5 from international users.
Cost via OpenRouter sits at approximately $1.40 per million input tokens and $4.40 per million output tokens. By comparison, GPT-5.5 is priced at $5 input / $30 output and Claude Opus 4.8 at $5 input / $25 output.
Built without Nvidia
Z.ai says GLM-5.2 was trained on roughly 100,000 Huawei Ascend 910B processors using Huawei’s MindSpore framework, with no Nvidia hardware at any stage of development. The achievement underlines China’s progress in building a domestic semiconductor stack capable of producing frontier-grade models.
ZCode and market reaction
Alongside the model, Z.ai launched ZCode on July 2 — an agentic coding harness designed to run autonomous, marathon-length engineering sessions. New users receive five million free tokens; existing subscribers get a 50% data quota increase.
Shares of Knowledge Atlas Technology, Z.ai’s Hong Kong-listed parent (HKEX: 2513), jumped as much as 48% following the announcement. The release continues a trend that began with DeepSeek earlier this year: the global share of Chinese AI models in developer usage has climbed from 3% to 13% in two months, according to industry tracking data.