GLM-5.3 tops open coding models, but Z.ai holds back the weights
The 743B model beats open rivals and undercuts US labs ~10x on price — but API and weights arrive in stages after its cyber scores tripped a safety review.
Copy markdown
Live today — but only inside the subscription
GLM-5.3 is available now through Z.ai's GLM Coding Plan (from ~$12.60/mo) and the ZCode platform. Standalone API and open weights are staged for release ~2 weeks out, once safety evals clear.
Best open, still chasing the US frontier
It scores 66.9% on DeepSWE (GitHub issue-fixing) and 34.5% on Z.ai's Code Bench (up from 5.2's 23.4%) — topping open-weight rivals but trailing Claude Fable 5 and GPT-5.6 Sol on coding and terminal tasks.
Why the weights are locked: it got too good at hacking
GLM-5.3 hits 84.5% on CyberGym and flagged 2,436 real vulnerabilities across 269 open-source projects, more than doubling 5.2's exploit scores. Z.ai says that capability is why raw weights sit behind a safety review.
Leaner per task, a fraction of US pricing
Built by scaling post-training on the 5.2 base, it finishes tasks in ~75K output tokens vs 5.2's 96K — beating Claude Opus 4.8 on token efficiency. Z.ai's line runs near a tenth of US frontier API rates.
Elsewhere: Claude now watermarks its text
Anthropic detailed how Claude embeds an invisible statistical watermark (plus C2PA metadata on images) marking output as processed by Claude. Worth knowing if you publish Claude-written copy where AI-detection matters.