Chinese artificial intelligence company Z.ai has released GLM-5.3, a new flagship model that it said can match or edge some leading US models in identifying software vulnerabilities while narrowing the gap with Anthropic on long-running coding tasks.

That comparison has an important limit. On CyberGym, which tests whether a model can inspect source code, find vulnerabilities, and verify them, Z.ai said GLM-5.3 scored 84.5%. That was slightly ahead of Anthropic’s Mythos 5 at 83.8% and OpenAI’s GPT-5.6 Sol at 83.6%. But Mythos 5 remains well ahead when the task shifts from finding a flaw to building a working exploit.

Z.ai said GLM-5.3 scored 54.4% on ExploitBench, compared with 78.0% for Mythos 5. In another timed test, GLM-5.3 completed 105 attack development tasks in two hours and 130 in six hours. Mythos 5 completed 181 and 247, respectively.

GLM-5.3 uses the same base model as its predecessor, GLM-5.2, with Z.ai attributing the improvement entirely to post-training, or the work performed after a foundation model completes its main pretraining run.

That makes GLM-5.3 less a jump to a larger foundation model and more a test of how far Z.ai can improve an existing one by training it in longer and more realistic software engineering environments.

There is also early independent evidence that the coding gains extend beyond Z.ai’s own tests. FrontierSWE, which tests models on difficult software engineering projects, currently ranks GLM-5.3 second in its tested Claude Code configuration, behind Claude Fable 5. GLM-5.2 ranks fifth. GLM-5.3 recorded an average rank of 4.50 and a 78% dominance score, compared with 2.88 and 88% for Fable 5.

The result does not make GLM-5.3 a stronger model than Fable 5 across the board. Rather, it suggests Z.ai is becoming more competitive with proprietary frontier systems in software engineering, long-horizon agent tasks, and some security-related coding work.

Z.ai’s own description makes that focus clear. In its launch blog, it emphasized GLM-5.3 as a model for “frontier coding with emergent cyber capabilities,” while co-founder Tang Jie described it on X as “built to code” and “ready for cyber defense.” Tang said the gains came from post-training on the same 743 billion-parameter base.

The company said GLM-5.3’s training goes beyond isolated programming exercises. Its training environments cover workflows that require the model to identify a problem, consider possible solutions, make changes, check the results, and deliver finished work. Some tasks, according to Z.ai, are designed to resemble several days of work by a senior engineer and require access to computing clusters, storage systems, internal documents, and code libraries.

That focus builds on GLM-5.2, which Z.ai released in June for long-running tasks. GLM-5.2 expanded the model’s context window to one million tokens and was designed to keep track of goals and engineering requirements over lengthy projects. GLM-5.3 keeps that context window and supports a maximum output of 128,000 tokens while remaining a text-focused model.

The gains are clear in Z.ai’s benchmark results. Its Terminal-Bench 3.0 score rose from 4.6 for GLM-5.2 to 28.3 for GLM-5.3. DeepSWE v1.1 increased from 46.2 to 66.9, while Agents’ Last Exam improved from 23.8 to 28.5. On Z.ai’s own Code Bench, the company claims a 50% improvement in coding performance over GLM-5.2.

The improvement was even larger in cybersecurity. GLM-5.3’s ExploitBench score more than doubled from GLM-5.2’s 24.4% to 54.4%, even though the base model did not change. Z.ai acknowledged that GLM-5.3 is currently better at the earlier stages of finding vulnerabilities, including code review, discovery, and verification, than at deeper exploitation or complete offensive and defensive operations.

That may explain why the company describes its cyber capabilities as “emergent.” GLM-5.3 was not trained as a dedicated cybersecurity model. Z.ai said these capabilities appeared as it expanded reinforcement learning and trained the model in longer and more varied software development environments.

Cybersecurity therefore serves as both a new selling point and a test of Z.ai’s broader idea: if a model is trained to work more like an autonomous software engineer, the same skills it uses to understand, debug, and change complex code may also help it find weaknesses in that code.

Z.ai has not yet released the model weights that would allow developers to download and run GLM-5.3 independently. The company said it would delay the public release for about two weeks while carrying out more security checks and strengthening safeguards. Its most sensitive cybersecurity functions will initially be limited to selected partners and verified users under a “trusted access” program.

Investors did not respond to the launch with another immediate rally. Z.ai’s Hong Kong-listed shares closed down 3.6% on August 14, the day GLM-5.3 was announced. That contrasted with the market enthusiasm surrounding GLM-5.2 earlier in the summer.

By late June, Z.ai’s shares had risen more than 2,000% from its January listing as GLM-5.2 drew attention for coming closer to leading US models in coding and agent tasks.

GLM-5.3 marks progress in a narrower set of capabilities. It does not show that Z.ai has overtaken Anthropic or OpenAI across frontier AI more broadly. Instead, it suggests the gap is narrowing in areas such as coding and cybersecurity, and that meaningful gains may not always require another large pretraining run.