Z.ai says GLM-5.3 edged Anthropic's Mythos 5 on a cyber test

Chinese AI firm Zhipu AI (HKG: 2513) has gone big-game hunting with the document it released along with the rollout of its latest GLM-5.3 model.

The firm, trading internationally as Z.ai, specifically named Anthropic’s Mythos 5 as one of the models that GLM-5.3 outperformed on the CyberGym benchmark, which measures a model’s capability to check source code for software vulnerabilities.

GLM-5.3 runs on the same base model as GLM-5.2, a 743-billion-parameter mixture-of-experts design that activates roughly 40 billion parameters per token.

How does Z.ai’s GLM-5.3 compare to other models?

According to the numbers popularized by Z.ai, its GLM-5.3 scored 84.5% on the CyberGym benchmark, which is less than a single point’s lead over Anthropic’s Mythos 5 (83.8%) and OpenAI’s GPT-5.6 Sol (83.6%).

Beating out Mythos could see Z.ai sliding into the number one spot for developers and security teams looking to deploy open-weight models to combat attacks. The firm has said that the model has flagged more than a thousand serious bugs in live software already.

On more difficult tests, such as ExploitBench and ExploitGym, Z.ai admits that its latest model trails the top US models.

For example, on ExploitBench, which asks a model to actually build a working exploit from vulnerabilities, GLM-5.3 scored 54.4% against Mythos 5’s 78%.

When will GLM-5.3 weights become available?

Z.ai has committed to releasing the weights of its new GLM-5.3 model to Hugging Face on August 28. This is the first GLM model weight to be held back, with the Chinese AI lab citing safety evaluation and hardening as reasons for the two-week delay.

More than 100 of the 2,436 vulnerabilities the model found were critically rated after reviewing about 269 projects, including software used in the Linux kernel, WinRAR, Redis and FFmpeg. The oldest flaw had reportedly gone unnoticed for 45 years.

Until then, only paying subscribers on the Z.ai Coding Plan or its ZCode harness can use it.

The numbers are Z.ai’s own

Every figure above comes from Z.ai’s announcement, not from an independent run. Even the public benchmarks were executed by Z.ai using its own configuration, so the internal Code Bench cannot be audited outside the company.

Independent leaderboards such as Artificial Analysis have not yet added the model as of this Cryptopolitan report.

The launch also lands against a live example of the stakes. Z.ai’s earlier GLM-5.2 was the model Hugging Face turned to after OpenAI test systems, run with reduced guardrails, escaped a sandbox and compromised its servers.

On the geopolitics, Zhipu struck a cooperative note, telling reporters that AI “should not be a solo performance by one nation, but a symphony of global collaboration.”

The smartest crypto minds already read our newsletter. Want in? Join them.