What happened

According to an announcement on X by Zhipu AI, the company released GLM-5.3 on its ZCode platform. The model achieved an 84.5 score on the CyberGym benchmark, outperforming Mythos 5 and GPT-5.6 Sol.

The announcement states that GLM-5.3 is trained on a 743B parameter base model and is built for cyber defense and coding tasks. It claims a dramatic improvement over GLM-5.2 with better results using fewer output tokens, and API access and open weights will be released in stages.

Evidence

• Zhipu AI released GLM-5.3 on ZCode with an 84.5 CyberGym score, above Mythos 5 and GPT-5.6 Sol.

  X · attributed

  Zhipu AI released GLM-5.3 on ZCode and it scored 84.5 points on CyberGym, scoring above Mythos 5 and GPT-5.6 Sol.

• GLM-5.3 is trained on a 743B base model and built for cyber defense and coding tasks.

  X · attributed

  Trained on top of a 743B base model. > Build for Cyber Defence and Coding tasks.

• GLM-5.3 delivers a dramatic improvement over GLM-5.2 with better results using fewer output tokens.

  X · attributed

  GLM-5.3 takes agentic coding to the next level, delivering a dramatic improvement over GLM-5.2 while achieving better results with fewer output tokens.

Why it matters

This release signals Zhipu AI's aggressive push into efficient agentic coding and cyber defense, potentially disrupting the competitive landscape among AI model providers.

Limits and uncertainties

The CyberGym score is reported by Zhipu AI and lacks independent verification.

The staged release of API access and open weights may delay access for many users.

Practical implications

Operators should monitor the staged release to plan integration, and the efficiency gains suggest potential cost savings in inference for coding workloads.

What to watch

The staged release of API access and open weights, which will allow independent benchmarking and wider adoption.

Independent evaluations on other coding benchmarks to validate CyberGym results.