GLM-5.3 is a post-training bump on the 5.2 base — Coding Plan only, API later, weights in two weeks

Z.ai published GLM-5.3 on August 14, 2026. It uses the same base model as GLM-5.2. Every reported gain comes from post-training. (Source: Z.ai blog, 2026-08-14)

Key facts:

• The official docs page says the GLM-5.3 API is coming soon. The live surface today is the GLM Coding Plan, plus ZCode. (Source: Z.ai GLM-5.3 docs)
• Weights are not public. Z.ai says it will release them “in two weeks after launch, once safety evaluation and hardening are complete.”
• Official context length is 1M tokens. Maximum output is 128K. Input and output are text.
• Terminal-Bench 3.0 is 28.3, up from 4.6 on GLM-5.2. DeepSWE v1.1 is 66.9, up from 46.2. Agents’ Last Exam is 28.5, up from 23.8.
• CyberGym is 84.5, ahead of Mythos 5 at 83.8 and GPT-5.6 Sol at 83.6. ExploitBench is 54.4, up from 24.4, and still well behind Mythos 5 at 78.0 and Sol at 76.5.

What this means if you’re building with GLM

1. “Open-weights SOTA” is a claim about a checkpoint you cannot download. The blog calls GLM-5.3 “the most capable open-weights model for coding” and “open-source SOTA” on Terminal-Bench 3.0 and Agents’ Last Exam. The same post says the weights ship in two weeks. Official docs still headline it as “#1 Open-Source Model.” Until the repo exists, the runnable open GLM is still GLM-5.2.

2. The useful number is token efficiency, not the crown. On Z.ai’s in-house Code Bench (Claude Code 2.1.207), GLM-5.3 at Max effort is 34.5% at roughly 75K output tokens per task, versus GLM-5.2 at 23.4% at 96K. At High effort it is 31.4% at around 50K, which the post says beats Claude Opus 4.8 at 29.5% with 120K. Claude Fable 5 still leads the same chart at 39.5% Max. That is a cost curve, not a frontier takeover.

3. Cyber is a discovery lead, not an exploit lead. CyberGym 84.5 is the SOTA line Z.ai emphasizes. ExploitBench 54.4 more than doubles 5.2’s 24.4 and still sits under Mythos 5 and GPT-5.6 Sol. Z.ai says the rest of the jump was unexpected. Treat that as a vendor narrative. Do not point this model at a production network because a blog says “emergent cyber.”

4. If you already pay for the Coding Plan, the clock is the price. Point usage is split across input, cached input and output. Off-peak calls use 50% of the standard points. Peak is 14:00–18:00 UTC+8, Monday through Friday. ZCode claims a 98%+ cache hit rate and a 1.5× quota boost through August 31. There is no public per-million-token API card on the launch page.

5. Do not migrate a self-hosted 5.2 box yet. Same base, new post-training, no weights, no public API id. Keep GLM-5.2 local as the open path. When the 5.3 checkpoint lands, that is the tutorial. Today’s move, if any, is an A/B inside the Coding Plan.

Sources: Z.ai — GLM-5.3 · Z.ai GLM-5.3 docs · GLM Coding Plan

Related: How to run GLM-5.2 locally · GLM-5.2 open weights · GLM-5.2 Semgrep cyber benchmark · Claude Code daily driver