Zhipu AI

Artificial Intelligence · Cybersecurity · Large Language Models

Launch

Zhipu launches GLM-5.3, beating Anthropic's Mythos 5 on cybersecurity test

Despite the benchmark win, investors sent shares down nearly 4% on concerns that agentic AI's rising inference costs are unsustainable for the company.

• Z.ai (Zhipu AI) released its flagship GLM-5.3 model on August 14, 2026, an open-weights system built on the same roughly 750-billion-parameter base as GLM-5.2, with full weights due to be published in about two weeks after security review.
• Beijing-based Z.ai builds the GLM line of coding and agentic models used with tools like Claude Code and OpenCode, and went public in Hong Kong in January 2026 at a market cap of nearly $7 billion.
• On CyberGym, a benchmark for finding and confirming code vulnerabilities, GLM-5.3 scored 84.5%, narrowly ahead of Anthropic's restricted Mythos 5 (83.8%) and OpenAI's GPT-5.6 Sol (83.6%), per Z.ai's self-reported results.
• GLM-5.3 lagged badly on ExploitBench, which measures turning found flaws into working attacks, scoring 54.4% versus Mythos 5's 78% and GPT-5.6 Sol's 76.5%, and completed fewer attack tasks in timed trials (105 vs. 181 in two hours).
• Working with Chinese security teams, Z.ai says GLM-5.3 identified 2,436 vulnerabilities across 269 real-world software projects, including flaws in code up to 40 years old, now logged in a public registry.
• The launch follows Z.ai's completion last month of a data center built on at least 10,000 Chinese-made chips, underscoring how domestic AI firms are pushing model capability despite US export controls on advanced chips.
• Z.ai shares fell nearly 4% on the day of the launch, with one analyst telling Bloomberg the company remains on an unsustainable commercial footing as agentic AI drives up inference costs and losses.
• The release shows Chinese open-weight models closing in on Western frontier systems specifically in vulnerability discovery, a dual-use capability that sharpens debate over whether such tools help defenders or arm attackers first.