Z.ai calls GLM-5.3 "the most capable open coding model," but flip open its own benchmark table and the story gets messier. On Terminal Bench 3.0, GLM-5.3 scored 28.3, trailing closed-source Claude Fable 5 (33.7) and GPT-5.6 Sol (34.6). On DeepSWE v1.1, which tests real GitHub issue resolution, its 66.9 also fell short of fellow open-source model Kimi K3's 67.5, as well as Fable 5's 69.7. An official document meant to prove GLM-5.3 is the best ended up making it pretty clear it isn't.
GLM-5.3 went live Thursday through the GLM Coding Plan subscription and ZCode, with API access and downloadable weights to follow in phases after security review. The model has 74.3 billion parameters, built on the GLM-5.2 base with expanded post-training. Z.ai's own words: "Scaling post-training is all we did for GLM-5.3" — over the past month, they added more environments, more diverse tasks, and more compute.
This release isn't about pushing scores higher — it's about saving tokens. On Z.ai's own Code Bench, GLM-5.3 hit 34.5% under Max effort, burning roughly 75,000 output tokens per task on average. GLM-5.2 scored 23.4% but burned 96,000 tokens. In other words: the new model scores higher while using less fuel. Compared to Claude Opus 4.8, Z.ai claims better token economics for GLM-5.3, but still admits it "remains behind Claude Fable 5, which reaches 39.5% at Max effort" — a line straight from their own blog.
The real highlight is in security testing. GLM-5.3 scored 84.5% on CyberGym, more than double GLM-5.2's score on vulnerability-discovery benchmarks. Z.ai says the model found 2,436 vulnerabilities across 269 open-source projects, with 1,097 rated medium-to-high risk. This is one of the few parts of the release that its own data didn't undercut.
Price is where open-source models can still compete. GLM-5.2's official API pricing runs $1.40 per million input tokens and $4.40 per million output tokens, with Z.ai claiming its overall API pricing is roughly one-tenth that of US frontier models. By comparison, GPT-5.3-Codex charges $1.75 in and $14 out, while Claude Opus 4.8 sits at the very top of Anthropic's price list. The GLM Coding Plan itself runs on a credit-based system, with off-peak usage discounted by half.
The "open-weights" label hasn't actually been fulfilled yet. The announcement states plainly that the weights won't be publicly released for another two weeks or so. What's available now is the subscription-based GLM Coding Plan and ZCode — not a downloadable file you can run yourself. Z.ai is a Beijing-based lab on the US Entity List, meaning US companies can't export controlled technology to it. Even so, Chinese open-source models have surpassed their American counterparts in token usage on OpenRouter, and the GLM series itself remains quite popular.
To sum up this self-undercutting announcement: GLM-5.3 beats its own predecessor and outperforms some open-source rivals, but on the most visible coding leaderboards, closed-source American frontier models still lead. And the "open-source" badge it's so proud of remains, for now, an IOU.
