China's Z.AI releases GLM-5.3, a 743-billion-parameter open-weight coding model

Chinese AI lab Z.ai releases GLM-5.3, a 743-billion-parameter open-weight coding model. Launched on August 13, 2026, the model scales post-training on its predecessor, GLM-5.2, to improve token efficiency and cybersecurity vulnerability detection. While GLM-5.3 outperforms open-source rivals like Kimi K3 on key benchmarks, it remains slightly behind leading closed-source U.S. models like Claude Fable 5. Despite being on the U.S. Entity List, Z.ai offers competitive pricing at a fraction of U.S. rates.

GLM-5.3 model release

• ▪Chinese AI lab Z.ai released its GLM-5.3 coding model on August 13, 2026, through its GLM Coding Plan subscription and ZCode platform.
• ▪Z.ai plans to release the API access and downloadable open weights for GLM-5.3 in stages following safety reviews, with weights scheduled for public release around late August 2026.
• ▪The GLM-5.3 model is a 743-billion-parameter coding model built by scaling post-training on the GLM-5.2 base model.

Coding benchmark performance

• ▪On the in-house Z.ai Code Bench at Max effort, GLM-5.3 achieved a score of 34.5%, outperforming GLM-5.2's score of 23.4%.
• ▪On the DeepSWE v1.1 benchmark, GLM-5.3 scored 66.9, trailing behind open-source rival Kimi K3 at 67.5 and closed-source model Claude Fable 5 at 69.7.
• ▪On the Terminal Bench 3.0 benchmark, GLM-5.3 scored 28.3, placing it slightly behind closed-source models Claude Fable 5 at 33.7 and GPT-5.6 Sol at 34.6.

Token efficiency improvements

• ▪The GLM-5.3 model outperforms the closed-source model Claude Opus 4.8 in terms of token economy on the Z.ai Code Bench.
• ▪The GLM-5.3 model focuses on token efficiency, consuming approximately 75,000 output tokens per task on the Z.ai Code Bench compared to GLM-5.2's 96,000 tokens.

Cybersecurity vulnerability detection

• ▪During testing, GLM-5.3 flagged 2,436 vulnerabilities across 269 open-source projects, identifying 1,097 of those vulnerabilities as medium-to-high severity.
• ▪The GLM-5.3 model achieved an 84.5% score on the CyberGym benchmark, more than doubling the exploitation benchmark performance of GLM-5.2.

API pricing structure

• ▪Z.ai prices its API at approximately one-tenth of United States frontier per-token rates, with GLM-5.2's official rate set at $1.40 input and $4.40 output per million tokens.
• ▪Z.ai's GLM Coding Plan operates on a points quota system where off-peak API calls cost half the standard rate.

U.S. export restrictions

• ▪The Beijing-based AI lab Z.ai is included on the United States Entity List, which prohibits American firms from exporting controlled technology to the organization.
• ▪Despite United States export restrictions, Chinese open-weight models have achieved high popularity, outperforming American models on OpenRouter token usage.