Z.ai released GLM-5.3 today, built on the same 743-billion-parameter base as GLM-5.2, with all improvements coming from scaled post-training alone. No new pretraining. The model jumps from 4.6 to 28.3 on Terminal-Bench 3.0, from 46.2 to 66.9 on DeepSWE v1.1, and from 26.2 to 48.2 on AutomationBench. It trails GPT-5.6 Sol and Claude Fable 5 on most benchmarks, but Z.ai's own Code Bench shows GLM-5.3 completing tasks at 34.5% accuracy using 75,000 output tokens where GLM-5.2 needed 96,000 to hit 23.4%. The efficiency gap matters more than the leaderboard position for enterprises running long-horizon coding agents.
The cybersecurity numbers are where the release gets complicated. GLM-5.3 scores 84.5% on CyberGym, edging GPT-5.6 Sol's 83.6%. On ExploitBench it hits 54.4%, more than double GLM-5.2's 24.4%. Z.ai says cyber capability progressed further along exploitation chains than expected as training scaled, an outcome the company describes plainly as faster than anticipated. Z.ai developer advocate Lou posted on X that GLM-5.3 already found a potentially serious vulnerability in Cursor, the AI coding tool recently acquired by SpaceX. Z.ai's own security work with Chinese teams has produced 2,436 vulnerability findings across 269 projects, with 1,097 rated critical or high severity and 2,383 still under embargo. The model is launching through Z.ai's GLM Coding Plan and ZCode environment only, with API access and open weights held back until safety evaluation and hardening are complete, approximately two weeks out.
Developers migrating existing GLM integrations cannot treat this as a drop-in swap. GLM-5.3 requires thinking to be enabled and will reject any request sending thinking.type: 'disabled'. Applications must update to specify a reasoning effort level before switching the model identifier or the call will fail. That is the operational detail buried in the release notes. The deeper question, the one worth reading the full technical announcement and Reuters reporting for, is what it means that a frontier-scale base model produced accelerating offensive security capability as a side effect of post-training scale, and how Z.ai's trusted access controls actually function in practice.
[READ ORIGINAL →]