GLM-5.3

Z.ai announced GLM-5.3 on 14 August 2026 with one methodological claim: "Scaling post-training is all we did for GLM-5.3." The base model is the same as GLM-5.2, so every reported gain comes from the post-training stack built in the previous cycle: IndexShare, SAO and slime. Terminal-Bench 3.0 moves from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9. Token efficiency rises at the same time: 34.5% on the private Z.ai Code Bench at roughly 75,000 output tokens in Max effort, against 23.4% at 96,000 for GLM-5.2, and ahead of Claude Opus 4.8 at 29.5% for 120,000. The claim of the most capable open-weights coding model does not follow from the published table: against Kimi K3 the scorecard is 3-3 with one tie.

The cyber result is what keeps the model in view. Z.ai calls the capability emergent while writing that it expected the training to produce it. CyberGym reaches 84.5% and ExploitBench 54.4%. Working with Chinese security teams, Z.ai reports 2,436 vulnerabilities identified across 269 open source projects, the oldest introduced in 1981; the Security Disclosure Ledger shows 53 disclosed and 2,383 under embargo.

Anthropic's Frontier Red Team, writing on 29 September 2026, counted 50 complete exploits in 410 ExploitBench attempts against 56 for Claude Mythos Preview, and found three guardrail bypasses that work: a fake red team framing (64% engagement), prefilled reasoning (92%), abliteration (100%). Z.ai promised the weights two weeks after launch, once safety evaluation and hardening were finished. That is a date, not a criterion.