CLAIM-2026-08-18-03

ExploitBench rose from 24.4% (GLM-5.2) to 54.4% (GLM-5.3).

Why it matters

The headline delta. Any practitioner treating GLM-5.2 evaluations as valid for GLM-5.3 is thirty points wrong on exactly the axis that matters for a coding agent's blast radius.

Sources

• z.ai/blog/glm-5.3
• MarkTechPost
• Unite.AI
• Kingy AI