Released on August 14, GLM-5.3 surprised even its creators with cybersecurity capabilities that emerged faster than expected. At the same time, the model delivers a significant improvement in coding performance. What makes this especially notable is that GLM-5.3 shares the same base model as GLM-5.2, with the improvements coming entirely from post-training.
Stronger Coding for Real Engineering Tasks
GLM-5.3 is trained on scenarios that mirror the real work of engineers, including tasks complex enough to represent several days of work for an experienced engineer. Rather than requiring users to break a large task into smaller steps and supervise every action, the model can take greater ownership of a task end-to-end.
The improvement is reflected in its coding performance. GLM-5.3 delivers a 50% improvement over GLM-5.2 on Z.ai’s in-house Code Bench and achieves open-source SOTA on public benchmarks including Terminal Bench 3.0 and Agents’ Last Exam.
More Coding Performance, Fewer Output Tokens
GLM-5.3 also improves token efficiency alongside coding performance. At Max effort, it reaches 34.5% at roughly 75K output tokens per task, compared with 23.4% at 96K for GLM-5.2. The same shift appears against closed models: at High effort, GLM-5.3 reaches 31.4% at around 50K output tokens, surpassing Claude Opus 4.8 at 29.5% with 120K.
The improvement therefore extends beyond coding performance alone: GLM-5.3 delivers stronger agentic coding results while consuming fewer output tokens.
An Emergent Capability in Cybersecurity
An unexpected capability in cybersecurity accompanies the coding improvements. GLM-5.3 reaches state-of-the-art performance on CyberGym for vulnerability discovery, with its largest gains appearing further up the exploitation chain, where it more than doubles GLM-5.2 on exploitation benchmarks.
The improvement is not limited to identifying isolated flaws. GLM-5.3 began to reason across multiple stages of exploitation, forming coherent plans for complete exploitation chains. Since GLM-5.2, when testing on real codebases with security teams, the model discovered 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues.
Now Available on FPT AI Factory
GLM-5.3 is now available on FPT AI Factory. You can request access today and run GLM-5.3 through a single endpoint, with the same API and billing experience as other models on the platform, including Qwen, MiniMax, and Gemini.
This makes it easy to compare and switch between models to optimize for cost and quality, while benefiting from straightforward, enterprise-ready billing with no international credit card requirement and technical support.
👉 Request access to GLM-5.3 today: https://short.factory.fpt.ai/OeQCO
Learn more about GLM-5.3 at https://z.ai/blog/glm-5.3
