Z.ai Claims GLM-5.3 Outperforms Anthropic's Mythos 5 on CyberGym Benchmark

Chinese AI firm Z.ai (HKG: 2513), trading internationally as Z.ai, has made a bold claim that its latest model, GLM-5.3, outperforms Anthropic's Mythos 5 on the CyberGym benchmark.

The CyberGym benchmark measures a model's capability to check source code for software vulnerabilities, and GLM-5.3 scored 84.5% on this test, just a single point ahead of Mythos 5's 83.8%.

Z.ai admits that its model trails top US models on more difficult tests, such as ExploitBench and ExploitGym, but the firm claims that GLM-5.3 has already flagged over a thousand serious bugs in live software.

The weights of the new model will be released to Hugging Face on August 28, but only paying subscribers can use it until then.