GLM-5.3: Frontier Coding with Emergent Cyber Capabilities

Benchmark Model Rank Results
agents-on-agents-last-exam-ale-cliGLM-5.3 (Claude Code)#3Pass Rate: 28.5
agents-on-posttrainbench-v1-1GLM-5.3 (Claude Code)#3Accuracy: 39.8
code-generation-on-terminal-bench-2-1GLM-5.3 (Claude Code)#4Accuracy: 88.2
exploit-generation-on-exploitbenchGLM-5.3 (Claude Code)#4Coverage: 0.544
vulnerability-reproduction-on-cybergymGLM-5.3 (Claude Code)#4Success Rate: 84.5
code-generation-on-nl2repoGLM-5.3#5Accuracy: 58.0
code-generation-on-terminal-bench-3-0GLM-5.3 (Claude Code)#5Accuracy: 28.3
agents-on-automationbench-publicGLM-5.3#6Success Rate: 48.2
agents-on-toolathlon-verified-108-tasksGLM-5.3#9Pass@1: 0.73
code-generation-on-programbenchGLM-5.3#10Almost Resolved: 19.0
code-generation-on-deepsweGLM-5.3 (MiniSWE-Agent)#13Pass@1: 0.669