Zhipu AI

Artificial Intelligence · Cybersecurity · Large Language Models

Launch

Z.ai's GLM-5.3-Flash nears Claude Opus 4.8 on coding at 1/10 the cost

The model was tested anonymously online before launch and became the week's most-used system on OpenRouter.

• Z.ai released GLM-5.3-Flash on August 26, 2026, its first natively multimodal GLM model, posting open weights to Hugging Face while pricing tasks at roughly $0.045 apiece.
• Z.ai is the commercial arm of Beijing-based Zhipu AI, one of a handful of Chinese labs, alongside DeepSeek, Alibaba's Qwen and Moonshot AI's Kimi, racing to ship frontier-class open models.
• The model packs 320 billion total parameters but activates only 18 billion per forward pass, combining sparse and linear attention with a roughly 1-million-token context window.
• On Z.ai's own coding benchmark at max reasoning effort, GLM-5.3-Flash scored 29.0 versus Claude Opus 4.8's 29.5, and it beat predecessor GLM-5.2 by wide margins, 63.4 to 46.2 on DeepSWE v1.1.
• Flash's launch followed the family's flagship GLM-5.3, whose open weights Z.ai delayed until August 28 after the model scored 84.5% on the CyberGym vulnerability-hunting benchmark, edging out Anthropic's restricted Mythos 5 at 83.8%.
• Open-weight models that can hunt software vulnerabilities nearly as well as access-restricted US systems raise the dual-use stakes just as Washington and Beijing spar over chip exports and model access.