Skip to main content

Zhipu's GLM-5.3: Same Size, 50% Smarter, Coding Skills Rival Claude

On August 14, Zhipu AI officially unveiled its latest large language model, GLM-5.3. The base parameter count remains at over 74 billion—the same as its predecessor, GLM-5.2—so the rumored jump to 1 trillion parameters didn't materialize. That upgrade might be saved for GLM-5.5. But don't let the unchanged size fool you: through clever post-training techniques, Zhipu managed to squeeze out a 50% performance boost over the previous generation, propelling GLM-5.3 to the top of several major open-source model benchmarks. Its programming and agent capabilities are now approaching those of Claude Fable5, and its coding experience surpasses other domestic models.

Benchmarks That Speak for Themselves

The numbers are hard to ignore. On Terminal-Bench3.0, which tests a model's ability to handle complex tasks in real terminal environments, the score skyrocketed from 4.6 to 28.3. On DeepSWE v1.1, a benchmark focused on long-range software engineering and continuous code modification, it jumped from 46.2 to 66.9. And on Agents' Last Exam, which covers real professional scenarios requiring cross-tool collaboration and long-range tasks, it climbed from 23.8 to 28.5. In GDPval-AA v2, which spans 44 professions and evaluates high-value knowledge work, GLM-5.3 scored 1769 points, showcasing its professional task execution based on programming skills. Overall, the model has made significant strides in complex software engineering, terminal operations, and broader real-world agent tasks.

Image

Efficiency Meets Effectiveness

What's particularly telling is Zhipu's self-developed Z.ai Code Bench—a set of evaluations that place the model in a real local development environment, executing end-to-end tasks under different thinking modes. It simulates, as closely as possible, the experience developers have when using a Coding Agent. The results show that GLM-5.3 has found a sweet spot between effectiveness and token usage. In High mode, it achieved an accuracy rate of 31.4%, surpassing Claude Opus4.8's maximum mode at 29.5%, while using only about 50,000 tokens per task on average. Opus4.8, by contrast, requires around 120,000 tokens. In other words, GLM-5.3 gets the job done with a much shorter execution path.

Image

From Benchmarks to Real-World Use

GLM-5.3 isn't just a paper tiger. It's already available on Zhipu's official programming tool ZCode and efficiency tool AutoClaw, and it's open to all users of the GLM Coding Plan and subscription users. Third-party coding platforms like TraeWork, TraeCode, Kousi, WorkBuddy, CodeBuddy, Qoder, QwenWork, CatPaw, JoyCode, and OpenCode have also opened early access. The API is set to launch soon, and the complete model weights will be open-sourced within two weeks, after necessary security reinforcement. Zhipu's strategy here is to limit the model's potential attack capability while retaining its defensive value. At 13:00 today, the quota for all GLM Coding Plan users was reset, and everyone can see their restored quotas in the backend usage statistics.

Key Points

  • GLM-5.3 keeps the same 740B parameter count but improves performance by 50% through post-training.
  • Significant benchmark gains: Terminal-Bench3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, Agents' Last Exam from 23.8 to 28.5.
  • Coding efficiency: 31.4% accuracy in High mode with ~50k tokens vs. Opus4.8's 29.5% with ~120k tokens.
  • Available now on Zhipu's tools and third-party platforms; API coming soon, open-source weights within two weeks.