Overview
- Z.ai, formerly known as Zhipu AI, launched GLM-5.3 on Thursday as part of its GLM Coding Plan and ZCode initiative.
- The lab claims it to be the "most advanced open-weights model for coding."
- GLM-5.3 features 743 billion parameters, developed by enhancing the GLM-5.2 base model.
The Chinese artificial intelligence lab Z.ai has introduced GLM-5.3, positioning it as the most powerful open-weights coding model currently available. This model is now accessible via the GLM Coding Plan subscription and ZCode, with API access and downloadable weights expected to follow a safety evaluation.
According to the company’s launch announcement, the development of GLM-5.3 primarily involved scaling up from GLM-5.2. "With GLM-5.2 we built the stack... Over the past month we kept scaling on this stack: more environments, more diverse tasks, and more compute spent training on them," they stated.
Myriad: When will OpenAI release GPT-6? Click to make your prediction.The focus of the development was on improving token efficiency rather than sheer performance. GLM-5.3 comprises 743 billion parameters and utilizes significantly fewer tokens per task compared to its predecessor. Parameters represent the factors a model manages during data processing, whereas tokens are the fundamental units of information that a model can handle or produce.
Z.ai reported that GLM-5.3 achieved a score of 34.5% on their internal Z.ai Code Bench at maximum effort, using around 75,000 output tokens per task, compared to GLM-5.2's 23.4% at 96,000 tokens. In comparisons with proprietary models, it surpasses Claude Opus 4.8 on token efficiency but still trails behind Claude Fable 5, which scored 39.5% at maximum effort.
When it comes to coding capabilities, GLM-5.3 shows strong performance, outperforming the Chinese model Kimi K3 on key benchmarks.
In the Terminal Bench 3.0 test, which evaluates autonomous shell and tool usage in real Linux environments, GLM-5.3 scored 28.3, slightly lower than closed models Fable 5 (33.7) and GPT-5.6 Sol (34.6). On the DeepSWE v1.1 benchmark, which assesses the ability to resolve real GitHub issues end-to-end, GLM-5.3 (66.9) was outperformed by Kimi K3 (67.5) and Fable 5 (69.7).
Overall, GLM-5.3 surpasses its predecessor and some open-source competitors, but closed American models continue to dominate the leading coding benchmarks.
In cybersecurity assessments, GLM-5.3 demonstrated significant advancements, achieving an 84.5% score on CyberGym and more than doubling GLM-5.2's performance on exploitation benchmarks. Z.ai noted that the model identified 2,436 vulnerabilities across 269 open-source projects, with 1,097 classified as medium-to-high severity.
According to Z.ai, "GLM-5.3 takes agentic coding to the next level, delivering a dramatic improvement over GLM-5.2 while achieving better results with fewer output tokens." The model is currently available via the GLM Coding Plan and ZCode, with API access and open weights to be released gradually after comprehensive safety assessments.
GLM-5.3 takes agentic coding to the next level, delivering a dramatic improvement over GLM-5.2 while achieving better results with fewer output tokens. pic.twitter.com/KGc6ZR7GHv
— Z.ai (@Zai_org) August 14, 2026
In terms of pricing, Z.ai’s open-weights model offers a more affordable alternative compared to U.S. frontier models. The GLM Coding Plan operates on a points quota system (with off-peak calls costing half), while Zhipu's API is priced at about a tenth of the rates for U.S. frontier models—GLM-5.2's official pricing was $1.40 for input and $4.40 for output per million tokens, compared to GPT-5.3-Codex at $1.75 for input and $14 for output, with Claude Opus 4.8 being among the higher tier prices from Anthropic. Z.ai, based in Beijing, is listed on the U.S. Entity List, restricting American companies from exporting controlled technologies to it. Nevertheless, the GLM model remains highly popular, with Chinese open-weight models already outperforming American counterparts in OpenRouter token utilization.
According to the launch announcement, the weights for GLM-5.3 are expected to be publicly released in approximately two weeks, indicating that the open-weights classification pertains to future releases rather than what is currently downloadable.