Z.ai GLM-5.3: Benchmark Results, Cyber Capabilities, and How It Stacks Up Against GPT-5.6 Sol and Mythos 5

The picture changes on ExploitBench, which requires deeper reasoning about how a real vulnerability could actually be exploited and demands a working exploit. GLM-5.3 more than doubles its predecessor's score, climbing from 24.4% to 54.4% . However, Z.ai reports that Mythos 5 scores 78% on ExploitBench and GPT-5.6 Sol scores 76.5% — leaving GLM-5.3 well behind on this more difficult measure .

On ExploitGym, which counts how many exploitation tasks a model can complete under a normalized two-hour budget, GLM-5.3 finishes 105 tasks (up from 29 for GLM-5.2) . Within a six-hour budget, that rises to 130 tasks .

Sources:

Real-World Vulnerability Detection

Beyond benchmarks, GLM-5.3 has been deployed in real-world security testing with impressive results:

• 2,436 confirmed vulnerabilities found across 269 open-source projects, with results tracked in a public Z.ai Security Disclosure Ledger .
• Of those, 1,097 were classified as critical or high-severity .
• The model found bugs in software dating back roughly 40 years .
• Shortly after launch, it also identified a serious vulnerability in Cursor, the AI-powered code editor .

Key Coding & General Benchmarks

GLM-5.3 also posts strong numbers on coding and agentic benchmarks, positioning it as a leading open-weight model for software engineering tasks:

• Terminal-Bench 3.0: scored 28.3 — highest among open-weight models and approaching Anthropic's Claude Fable 5 .
• DeepSWE v1.1: jumped from 46.2 to 66.9 .
• Z.ai internal Code Bench (Max effort): 34.5% (vs. GLM-5.2's 23.4%), using roughly 75,000 output tokens per task versus 96,000 .
• Agents' Last Exam: first place among open-weight models .
• World of AI Benchmark composite: ranked #6 overall with an 81.3 composite score, just behind Kimi K3 (81.5) and ahead of GPT-5.6 Terra .
• FrontierSWE: reached 78.1, up from 67.5 .
• SWE-Marathon: doubled from 19.4 to 42.5 .

Why Open Weights Are Delayed

Z.ai says the model's cybersecurity capability "grew faster than anticipated" during post-training . The base model's 740 billion parameters were not retrained — all gains came from post-training improvements . Because the capability was an emergent surprise, the company is holding back the open-weight release for two weeks while it strengthens safety and security controls . The model is already available via the Z.ai API and through integrations with ZCode, Claude Code, and OpenCode .

Important Caveats

• All GLM-5.3 benchmark scores are company-reported and have not yet been independently verified .
• The CyberGym advantage over Mythos 5 and GPT-5.6 Sol is narrow — less than one percentage point .
• On ExploitBench, GLM-5.3 trails US rivals by a wide margin (54.4% vs. 78% for Mythos 5 and 76.5% for GPT-5.6 Sol) .
• For context, a broader cybersecurity benchmark aggregator lists GPT-5.6 Sol at 84.5% on CyberGym (matching GLM-5.3) and Mythos 5 at 83.8%, while showing an ExploitBench of 73.5% for Sol and 78% for Mythos 5 .