The most consequential thing Z.ai has done is not a model release, it is a pricing decision. The GLM Coding Plan gives monthly access to their models through the tools people already use for agentic coding, explicitly including Claude Code and Cline, starting around eighteen dollars a month. Set that against what equivalent usage costs on frontier American models and the gap is not incremental. For individual developers and small teams paying for their own tooling, that is a genuinely disruptive number and it explains why the plan has attracted the attention it has.
The documentation is built to make that switch easy. The model lineup is documented around GLM-5.3 as the flagship, positioned on software engineering and agent capability, GLM-5.3-Flash as a multimodal coding model oriented to visual programming, GLM-Image for text-to-image, and CogVideoX for video generation. Code examples run through cURL, Python, Java and JavaScript, there are official SDKs for Python and Java with public repositories, and OpenAI SDK compatibility is documented with examples, which is what actually matters for adoption because it turns a migration into a configuration change. There is an llms.txt index at docs.z.ai for feeding the whole reference to a model.
The reference quality is good. Parameters are described rather than merely listed, the API reference is complete, and the examples run. For a documentation set that many readers will arrive at from a different language and culture of technical writing, the English is clean and the structure is conventional in the ways that matter. Two things are underdocumented and both matter.
The first is the limits on the coding plan. A subscription that gives you agentic coding for eighteen dollars a month has to have constraints somewhere, whether that is rate limits, concurrency, fair-use thresholds or model access, and those constraints are not prominent. Finding out where the ceiling is by hitting it mid-task is a bad way to learn. The second, and larger, is data.
There is no meaningful treatment in the developer documentation of what happens to the code you send, where inference is performed, what is retained or for how long. Agentic coding tools see your entire repository. For a personal project that is a question you can wave away. For an employer's codebase it is not, and the answer needs to come from a policy document rather than from an absence of discussion.
Zhipu is a Chinese company and for many organisations that fact alone determines the answer regardless of what the retention policy says. I want to be even-handed about this: the same question applies to every provider and most teams never ask it of the American ones either. The difference is that the American labs publish enterprise data commitments prominently and Z.ai's developer documentation does not. That is a documentation gap rather than evidence of a problem, and it is a gap a buyer has to close themselves.
Model naming also moves fast enough that anything written more than a couple of months ago will reference versions that have been superseded. Trust the docs, not the blog posts. Three point eight. Clear, complete, well-exampled documentation for the most aggressive pricing in agentic coding, marked down for silence on data handling and for plan limits that should be far more prominent than they are.