GLM-5.3: Frontier Coding with Emergent Cyber Capabilities

Read https://z.ai/blog/glm-5.3 for the benchmark figures on models/glm-5-3. As recorded by the reader: GLM-5.3 publishes only one allowed benchmark: the card/blog table labels it 'Terminal Bench 2.1' (score 88.2, GLM-5.3 column; corroborated by the identical row in the z.ai release blog). Mapped to canonical 'Terminal-Bench 2.1'. Everything else in that table is outside the allowed list (Terminal Bench 3.0 — different version, excluded; DeepSWE v1.1, NL2Repo, ProgramBench, FrontierSWE, SWE-Marathon, PostTrainBench, CyberGym, ExploitGym, ExploitBench, Toolathlon Verified, AutomationBench, Agents' Last Exam, HLE w/ Tools, GDPval-AA v2). The card and blog publish NO GPQA Diamond, AIME, MMLU/MMLU-Pro, SWE-bench Verified/Pro, LiveCodeBench, HumanEval, MATH-500, IFEval or Arena-Hard score — because GLM-5.3 is a post-training-only release over the GLM-5.2 base, its published table focuses on coding/cyber/agentic evals. Reasoning effort defaults to 'max'. Not to be confused with GLM-5.3-Flash.