Z.ai Launches GLM-5.3 With Major Coding and Cybersecurity Improvements

Chinese AI company Z.ai has released GLM-5.3, the latest model in its increasingly powerful and largely open-weight GLM language model family. The new AI model delivers substantial improvements in long-horizon coding, software engineering, and cybersecurity research.

GLM-5.3 has already demonstrated advanced cybersecurity capabilities, including the discovery of potentially serious vulnerabilities in AI coding startup Cursor. Cursor was acquired by SpaceX, according to Z.ai developer advocate Lou in a post on X. VentureBeat has contacted Cursor for confirmation and is awaiting a response.

For now, GLM-5.3 is available only through Z.ai’s GLM Coding Plan and ZCode Coding Environment. The company says API access and model weights will be released later, “once safety evaluation and hardening is complete.” Z.ai expects to release the weights approximately two weeks after the initial launch.

The release is significant for enterprise developers because it demonstrates how far an AI model can improve through post-training alone. Z.ai says GLM-5.3 uses the same fundamental base model as GLM-5.2. Its performance gains come from scaling post-training across more environments, a wider range of tasks, and additional reinforcement learning compute.

In other words, GLM-5.3 is an experiment in extending the capabilities of an existing foundation model without beginning another costly pre-training cycle.

“Post-training scaling is all we did for GLM-5.3,” Z.ai writes in its technical announcement.

The results indicate that there is still significant room for improvement after a model’s initial training. However, Z.ai also encountered an important challenge: cybersecurity capabilities improved faster than expected as training scaled, particularly when tasks progressed from finding vulnerabilities to developing complete exploit chains.

Reuters reported on Friday that Z.ai is introducing additional controls for some of GLM-5.3’s most advanced capabilities, including a “trusted access” system for sensitive functions.

GLM-5.3 Delivers a Major Coding Performance Increase Without a New Base Model

Instead of replacing GLM-5.2, GLM-5.3 builds on the same large-scale base model, which contains approximately 743 billion parameters. Z.ai expanded the post-training system developed around GLM-5.2, with a particular focus on long-term reinforcement learning and autonomous software engineering.

These training environments are designed to resemble real engineering assignments rather than short, isolated programming tests.

According to Z.ai, AI agents can receive access to codebases, technical documentation, compute clusters, storage systems, and experiment results. They must then diagnose problems, modify software, run tests, evaluate results, and demonstrate measurable improvements while maintaining accuracy. Some assignments are designed to take an experienced engineer several days to complete.

This approach has produced sharp improvements across several coding and agent benchmarks.

GLM-5.3 improves its score on Terminal-Bench 3.0 from 4.6 to 28.3. On DeepSWE v1.1, the model increases its score from 46.2 to 66.9, while its AutomationBench score rises from 26.2 to 48.2. Agent Final Exam CLI also improves from 23.8 to 28.5.

GLM-5.3 does not outperform every competing frontier model. Z.ai’s benchmark table reports a Terminal-Bench 3.0 score of 34.6 for GPT-5.6 Sol and 33.7 for Claude Fable 5, compared with 28.3 for GLM-5.3. On DeepSWE v1.1, GLM-5.3 scores 66.9, while GPT-5.6 Sol scores 72.7 and Fable 5 scores 69.7.

However, benchmark position is only one part of the story. Z.ai is also emphasizing efficiency, including the number of output tokens required to complete complex tasks.

On Z.ai’s private Code Bench evaluation, GLM-5.3 achieves a 34.5% success rate using maximum inference settings and approximately 75,000 output tokens per task. GLM-5.2 achieves 23.4% while using approximately 96,000 output tokens. At high effort, GLM-5.3 reaches 31.4% with around 50,000 output tokens. Z.ai reports that Claude Opus 4.8 achieved 29.5% while using approximately 120,000 tokens.

Because Code Bench is a private, company-developed evaluation, these comparisons should be viewed as Z.ai-reported results rather than independent measurements. Even so, completing more tasks with fewer tokens could be highly valuable for businesses deploying AI coding agents. Inference costs and response times can increase quickly when agents run through lengthy reasoning, testing, and correction loops.

GLM-5.3’s Cybersecurity Capabilities Improved Faster Than Expected

Cybersecurity is one of the most notable developments in the GLM-5.3 release.

Z.ai added a vulnerability discovery environment to the model’s post-training process to improve its ability to identify software defects. The company says the model progressed beyond vulnerability discovery and began advancing more rapidly through the broader exploitation process.

“As we scaled up post-training, our cyber capabilities developed faster than expected,” Z.ai wrote.

On CyberGym, a benchmark that evaluates vulnerability discovery and validation in source code, GLM-5.3 scores 84.5%, compared with 77.2% for GLM-5.2. Z.ai reports that this result is also higher than the scores of 83.6% for GPT-5.6 Sol and 83.8% for Mythos 5.

The model’s advantage is less pronounced across the full exploitation stack. GLM-5.3 scores 54.4% on ExploitBench, more than twice GLM-5.2’s 24.4% score. However, it remains behind the scores reported by Z.ai for GPT-5.6 Sol at 76.5% and Mythos 5 at 78%.

On ExploitGym, GLM-5.3 completes 105 tasks within a normalized two-hour budget and 130 tasks within six hours. GLM-5.2 completes 29 and 39 tasks under the same conditions. Fable 5 completes 181 and 247 tasks, while GPT-5.6 Sol completes 216 and 293.

Although GLM-5.3 does not lead every cybersecurity benchmark, the direction of its progress may be more important than its position on the leaderboard.

Z.ai says that, working with Chinese cybersecurity teams, the model helped identify 2,436 vulnerabilities across 269 projects through expert review, screening, and deduplication. The company’s disclosure ledger lists 1,097 vulnerabilities as critical or high severity. Of those, 53 have been publicly disclosed, while 2,383 remained under embargo at the time of publication.

These results highlight a growing challenge for AI developers. The same long-running agent capabilities that make models more useful for software engineering can also make them more effective at security research—and potentially more capable of supporting cyberattacks.

GLM-5.3 Requires Changes to Existing API Integrations

Developers migrating existing GLM applications should be aware that GLM-5.3 introduces incompatible API behavior.

The model supports three inference-effort levels: low, high, and max. Z.ai recommends the max setting for coding tasks. Unlike earlier GLM releases, however, GLM-5.3 does not allow developers to disable its reasoning process.

Applications that currently send thinking.type: "disabled" must change the value to enabled and specify the desired inference effort before switching to the GLM-5.3 model identifier. Z.ai warns that requests may fail if this configuration is not updated.

As a result, adopting GLM-5.3 may require a genuine application migration rather than simply replacing an older model name in production code.

From GLM-4.5 to GLM-5.3: Z.ai’s Push Into Agent Engineering

GLM-5.3 is the latest stage in Z.ai’s rapid shift toward coding agents, autonomous software development, and long-running engineering workflows. The company was previously known as Zhipu AI.

GLM-4.5, released in July 2025, established much of this strategy. The 355-billion-parameter mixture-of-experts model was designed for reasoning, coding, and agent tasks. Z.ai also released the smaller 106-billion-parameter GLM-4.5-Air and promoted integrations with popular agent frameworks.

GLM-4.6 followed in September, expanding the context window from 128,000 tokens to 200,000 tokens. The model targeted coding, tool use, and agent workflows in environments including Claude Code, Cline, Roo Code, and Kilo Code. Z.ai also began placing greater emphasis on token efficiency in practical coding tasks rather than relying solely on benchmark scores.

A larger architectural change arrived with GLM-5 in February 2026. Z.ai increased the model from 355 billion parameters in GLM-4.5 to approximately 744 billion, with 40 billion active parameters. The company also expanded pre-training data to 28.5 trillion tokens, introduced its “Slime” asynchronous reinforcement learning infrastructure, and repositioned the GLM family around long-term “agent engineering” tasks.

By June, GLM-5.2 had turned that technical strategy into a clearer enterprise proposition. The 753-billion-parameter model offered a stable one-million-token context window, open weights under the MIT license, and compatibility with more than 20 coding environments. Z.ai also introduced IndexShare, a technique that reuses indexers across sparse attention layers to reduce the computational cost of processing very long contexts.

GLM-5.2 was priced at $1.40 per million API input tokens and $4.40 per million output tokens, with lower rates for cached inputs. That pricing positioned Z.ai as both a technical and cost-focused competitor to leading frontier AI labs.

Z.ai’s ambitions extend beyond model development. Reuters reported last month that Zhipu AI planned to raise approximately HK$31.4 billion, or about $4 billion, through a Hong Kong stock offering. The company said the funds would support research and development, computing infrastructure, hiring, and business expansion.

Together, the company’s recent releases show a consistent progression. GLM-4.5 combined reasoning, coding, and agent functions. GLM-5 significantly increased the scale of the underlying model. GLM-5.2 focused on long-context processing and long-horizon engineering. GLM-5.3 now seeks to extract additional performance from that same foundation through post-training.

GLM-5.3 Pricing, ZCode, and Availability

GLM-5.3 is currently available through Z.ai’s GLM Coding Plan and the company’s ZCode coding environment.

ZCode is Z.ai’s proprietary coding agent platform. It supports long-running “goal” tasks that allow an agent to plan, implement, test, and validate software projects. The platform also provides remote controls for active tasks and is available on macOS, Windows, and Linux.

Individual GLM coding plans currently begin at a promotional price of $12.60 per month for the Lite tier, which includes 10,000 credits per week. The Pro plan is listed at $56 per month and includes six Lite uses. The Max plan costs $117.60 per month and includes 14 Lite uses. Team Standard and Premium seats cost $88 and $188 per user per month, respectively.

Z.ai has also moved its coding plans to a points-based quota system. The system calculates input tokens, cached input tokens, and output tokens separately. Requests made during weekday off-peak hours cost 50% of the standard number of points.

The company has not yet announced general API pricing for GLM-5.3 in the available launch materials. Direct cost comparisons with GLM-5.2 and competing frontier models will therefore remain difficult until broader API access and tiered pricing are introduced.

This controlled rollout could ultimately become one of the most important aspects of the GLM-5.3 launch.

Over the past year, Z.ai has pursued an open model strategy based on permissive licensing, lower-cost inference, and compatibility with the existing coding-agent ecosystem. GLM-5.3 demonstrates what can happen when that strategy produces particularly rapid gains in a sensitive area. More capable autonomous engineering agents can also become more capable autonomous cybersecurity researchers.

The result is a model that advances Z.ai’s coding ambitions while forcing the company to confront the same capability and access trade-offs faced by the largest closed AI laboratories.

For enterprise developers, GLM-5.3 is notable for two reasons. First, its coding performance suggests that more capable agents can emerge through improved post-training and better task environments without continuously rebuilding the underlying model. Second, its cybersecurity results show why deployment controls and access policies are just as important as the methods used to train advanced AI systems.