GLM-5.3: benchmark results

Zhipu's GLM-5.3 (August 14, 2026), an open-weight coding- and cybersecurity-focused update built on the GLM-5.2 base with additional post-training. Provider: Zhipu. Released 2026-08-14. Access: Open.

Aggregate Rating 1836 ± 5, rank #39 of 1656 rated models, from 774 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
Arabic Broad Leaderboard - RAG QA9.73Average Score (0-10)100
EC2-Bench - Completion82.7Task completion score (0-100; weighted rule, environment-sta100
ExploitGym (version unspecified; rescaled 2h timeout)105Successful intended exploits (#); version unspecified; resca100
ExploitGym (version unspecified; rescaled 6h timeout)130Successful intended exploits (#); version unspecified; resca100
ProcArena - PostgreSQL - Interactive - Document-Grounded Synthesis73.9Execution Accuracy (%)100
ProcArena - PostgreSQL - Interactive - From-Scratch Synthesis53.5Execution Accuracy (%)100
Reflection Beam Preview 2026-10-05 - AA Omniscience (index) - public split only (our results)22.6Index (-100 to100)100
Reflection Beam Preview 2026-10-05 - SciCode59Score (%)100
Reflection Beam Safety - Irregular CTF - Group - Atomic77.5Accuracy (%)100
Reflection Beam Safety - WMDP-Chem - Log-likelihood78.2Accuracy (%)100
Robo Use - Agents - BEHAVIOR-1K100Solved / scored tasks (%)100
SWE-Milestone58.75Milestone Score (%)100

Interactive version: theaggregate.ai/model?slug=glm-5-3 · How It Works · Data refreshed daily, snapshot 2026-10-20.