GLM-5.3 Ships to Coding Plan Subscribers, Weights Held Back

Z.ai released GLM-5.3 on 14 August 2026, initially to GLM Coding Plan subscribers and through ZCode. The company's own documentation lists API access as "available soon" rather than live, and the public pricing table still shows GLM-5.2 at $1.4 input / $4.4 output per million tokens with no GLM-5.3 row - so there is no official GLM-5.3 API price to quote yet.

The model ID is glm-5.3. Z.ai documents a 1M-token context window, a 128K maximum output length, text-only input, and reasoning that is always on with three selectable thinking-effort levels (low, high, max). Function calling, context caching, structured JSON output and streaming are all supported.

GLM-5.3 runs on the same base model as GLM-5.2 - no new pre-training run. Z.ai attributes the entire improvement to scaled post-training: more task environments, more environment types, and longer training.

Vendor-reported benchmark deltas

These are Z.ai's own published figures for GLM-5.3 against GLM-5.2, not independent measurements:

On Z.ai's internal Code Bench the model scores 34.5% at max thinking effort (around 75K tokens of reasoning) and 31.4% at high effort (around 50K tokens).

Where GLM-5.3 loses

Several outlets framed this release as a Chinese model overtaking Western labs on cybersecurity. Z.ai's own numbers only partly support that. GLM-5.3's CyberGym result of 84.5% does edge Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). But on ExploitBench it scores 54.4% against Mythos 5's 78.0%, and on the ExploitGym task-completion test it finishes 105 tasks in two hours and 130 in six, against Mythos 5's 181 and 247. On two of the three cyber evaluations Z.ai published, GLM-5.3 is behind.

Weights

GLM-5.2 weights are on GitHub under Apache 2.0 at zai-org/GLM-5, which documents the shared base as 744B total parameters with 40B active per token. GLM-5.3 weights were not published at launch; coverage at the time reported them as expected roughly two weeks out. Until they appear, GLM-5.3 is effectively a subscription-only model despite sitting on an open-weight lineage.

Note on parameter counts: much of the third-party coverage cites 743B. The figure above (744B total / 40B active) is the one in Z.ai's own repository for the shared base model.