GLM-5.3 is an open-weight model from Zhipu AI (Z.ai), released August 29, 2026, with the company claiming strong benchmark positions against Claude and GPT. Anthropic's September 29 evaluation found that GLM-5.3 can autonomously exploit software vulnerabilities close to the level of Anthropic's own Claude Mythos Preview, and that its refusal mechanism can be bypassed with little effort.
Z.ai released GLM-5.3 as an open model on August 29, 2026, at 0:00 Japan time, following a promise made on August 14. The model uses a mixture-of-experts architecture with 744 billion total parameters and 40 billion active parameters. Third-party organization Artificial Analysis gave it an Intelligence Index score of 60 and an Agentic Index score of 59, which the company presented as surpassing Claude Fable 5 and GPT-5.6 Sol in third-party tests.
In a September 18 blog post, Z.ai described the production inference service for GLM-5.3-Flash running on a cluster of more than 100,000 Chinese-made AI accelerators. The company said most of the infrastructure work was carried out by an infrastructure agent powered by GLM-5.3, and that optimizations brought end-to-end service performance to about 3x with cost per token in line with mainstream NVIDIA GPUs.
On September 29, Anthropic published an evaluation finding that GLM-5.3 can autonomously find software vulnerabilities and build working attack code at a level close to Anthropic's Claude Mythos Preview. On ExploitBench, which tests exploitation of a known flaw in the V8 JavaScript engine used by Google Chrome, GLM-5.3 succeeded 50 times out of 410 attempts (about 12 percent) against 56 out of 410 (about 14 percent) for Claude Mythos Preview. On a control-flow hijacking benchmark over 100 tasks, GLM-5.3 scored 4 percent and Claude Mythos Preview 6 percent, while Kimi K3, DeepSeek-V4.1-Flash, Claude Opus 4.6 and GLM-5.2 all scored zero. Anthropic said the model's refusal mechanism can be bypassed with little effort, and that a version with weakened safety features appeared from a third party within days of release.
Synthesized by Yomimono from the cited Yomimono stories below, each itself
sourced, then editorially reviewed. Every
fact links the story it came from.
9h ago
Z.ai announced GLM-Legal, a model tuned on GLM-5.3-Flash for legal work. Its harness is also optimized for legal tasks. The company compared benchmark scores against GLM-5.3-Flash, Kimi K3, Claude Fable 5, and GPT-6 Astra, saying GLM-Legal improved on its base model and surpassed Claude Fable 5 and GPT-6 Astra on some tests. GLM-Legal is in beta testing, is scheduled to reach Z.ai's AI agent AutoClaw, and is accepting API access applications.
Oct 3
Anthropic published an evaluation on September 29 finding that GLM-5.3, the open-weight model from Chinese developer Zhipu AI (Z.ai), can autonomously find software vulnerabilities and build working attack code at a level close to Anthropic's own Claude Mythos Preview. On ExploitBench, which tests exploitation of a known flaw in the V8 JavaScript engine used by Google Chrome, GLM-5.3 succeeded 50 times out of 410 attempts, about 12 percent, against 56 out of 410, about 14 percent, for Claude Mythos Preview. On a control-flow hijacking benchmark over 100 tasks, GLM-5.3 scored 4 percent and Claude Mythos Preview 6 percent, while Kimi K3, DeepSeek-V4.1-Flash, Claude Opus 4.6 and GLM-5.2 all scored zero. Anthropic says the model's refusal mechanism can be bypassed with little effort, and that a version with weakened safety features appeared from a third party within days of release.
Sep 18
Z.ai explained in a blog post how it built the inference infrastructure behind the production service for GLM-5.3-Flash, running on a cluster of more than 100,000 Chinese-made AI accelerators. The company says most of the work was carried out by an infrastructure agent powered by GLM-5.3, and that optimizations brought end-to-end service performance to about 3x and cost per token in line with mainstream NVIDIA GPUs.
Aug 29
Z.ai released its GLM-5.3 model as an open model on August 29, 2026, at 0:00 Japan time. The MoE model has 744 billion total parameters and 40 billion active parameters. Third-party organization Artificial Analysis gave it an Intelligence Index score of 60 and an Agentic Index score of 59.