Zhipu AI released GLM-5.3 on August 14 and said the new model delivered a 50% improvement in its internal coding evaluations over GLM-5.2. At 13:00 that day, the company also reset GLM Coding Plan quotas for all users, making the launch immediately visible to anyone using the service.
The timing matters because GLM-5.3 is not a fresh base model. Zhipu AI said it kept the same foundation as GLM-5.2 and pushed the gains through post-training, a path that can move performance quickly without rebuilding the entire model from scratch. That is why the company is framing the release as a major step in coding and agent work rather than a routine update.
On the numbers, the jump is large. GLM-5.3 scored 28.3 on Terminal Bench 3.0, up from 4.6 for GLM-5.2, and 66.9 on DeepSWE v1.1, compared with 46.2 for the earlier model. It also reached 28.5 on Agents' Last Exam and 1,769 points on GDPval-AA v2, giving Zhipu AI a set of results it says puts the model first among open-source systems on Terminal Bench 3.0 and Agents' Last Exam (CLI).
The company is also leaning hard on cybersecurity benchmarks. On CyberGym, GLM-5.3 scored 84.5%, ahead of GLM-5.2's 77.2%, and slightly above Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%. But the picture is less uniform elsewhere: on ExploitBench, GLM-5.3 reached 54.4%, a clear improvement on GLM-5.2's 24.4%, yet still trailed Mythos 5 at 78.0% and GPT-5.6 Sol at 76.5%.
That split matters because it shows where the new release is strongest and where it still has ground to cover. On ExploitGym, GLM-5.3 completed 105 tasks within two hours and 130 within six hours, up from 29 and 39 for GLM-5.2, while Mythos 5 completed 181 and 247 tasks over the same windows. Zhipu AI says the work was driven by reinforcement learning on the same base model and built on IndexShare, SAO and the continuously evolving next-generation Slime framework, which helps explain the gains without changing the core model.
Zhipu AI says GLM-5.3 is now its strongest open-source coding model and that its coding and agent capabilities are approaching Claude Fable 5. Even so, the company also acknowledged limits in its own framing, saying, “We may still be far from reaching the intellig.” The next marker is already set: the open-source weights are due in two weeks, and that release will show how much of this performance can be carried beyond the company’s own benchmarking setup.
