Overview
GLM-5.3 is zhipu-ai / Z.ai’s Aug 2026 coding-focused model update sharing the same 743B MoE base as glm-5-2, with gains attributed entirely to scaled post-training. Open weights released August 28, 2026.
Recent Developments
- 2026-09-29: anthropic research + caisi assessment — most cyber-capable open-weight model to date; matches claude-mythos Preview on end-to-end exploits but safeguards bypassable 64–100% (2026-09-29-anthropic-glm-5-3-cyber-capabilities, anthropic-glm-5-3-cyber-capabilities)
- 2026-09-17: NIST CAISI — lags US frontier ~4 months on aggregate cyber benchmarks; SEC-Bench Pro 40.4%, ExploitBench 61.1% (2026-09-29-anthropic-glm-5-3-nist-caisi)
- 2026-08-13/14: Launch — vendor CyberGym 84.5%; ExploitBench 54.4%; 2,436 vulns / 1,097 critical-high claimed; open-weight-safety-delay ~Aug 28
- Thinking layer cannot be disabled on API (per TechTimes coverage)
- Integrates with ZCode, claude-code, OpenCode via GLM Coding Plan
All cyber metrics and vuln ledgers are company-reported pending independent audit.