Z.ai reports 1 million ZCode users and resets GLM plan limits
The Beijing AI company reset GLM Coding Plan limits, but its self-reported ZCode milestone does not distinguish registered, active or paying users.
By Ryan Merket · Published
Why it matters
ZCode gives Z.ai a direct route from its GLM models to paid developer usage. Subscription conversion, completed tasks and the inference cost of long-running agents will determine whether the product can support a durable coding business.
Z.ai, the Beijing AI developer founded by Tsinghua University researchers Jie Tang, Li Juanzi and Zhang Peng, is using its ZCode desktop coding application to build a developer business around its GLM models. Z.ai's official account says ZCode has reached 1 million users and that Z.ai reset usage limits for all GLM Coding Plan users.
The post does not provide evidence for the user count. The supplied material does not establish whether the 1 million figure counts registered, active or paying users, and it does not establish the announcement's publication date. The post provided no ZCode revenue, retention or paid-conversion figures.
The quota reset gives developers who had exhausted their allocations another chance to run repository-scale workflows. It could also increase inference demand on Z.ai as customers use agents to read repositories, execute commands, inspect browser output and revise work across multiple steps.
Z.ai is building distribution around GLM
ZCode is the product layer in Z.ai's effort to turn GLM usage into a recurring developer relationship. The GLM Coding Plan works with more than 20 coding tools, including Claude Code and OpenClaw. ZCode puts that model access inside Z.ai's branded interface with task history and controls for multi-step work.
RuntimeWire's July coverage of ZCode examined Z.ai's use of a free desktop application, lower-cost Coding Plan subscriptions and remote control through mobile and messaging apps. Z.ai advertises subscriptions starting at $16.20 a month, while the ZCode product page lists Lite, Pro and Max tiers. Higher-priced plans offer larger allowances and faster generation.
The ZCode documentation identifies GLM-5.2 as the underlying model and specifies a 1-million-token context window. It describes workspace state, file references, command execution, Git context, browser automation, scheduled work and remote control through mobile, Feishu and WeChat. Those capabilities place ZCode in the market for agents that can continue multi-step engineering work instead of stopping after suggesting an edit.
The update focuses on keeping agents on task
Z.ai described the update in its X post as a way to turn "long-horizon capabilities into completed work." A large context window lets a model read more of a repository, while the runtime still has to preserve the objective, verify changes and determine when the work is complete.
ZCode's Goal Mode documentation describes a control layer that stores the objective, status, budget, verification results and next action outside the model's narrative. The agent can continue across iterations while the runtime tracks whether the work has satisfied the stated goal. That structure is intended to make long tasks recoverable and auditable when a session fails, a model changes or a user returns later.
Separately, ZCode's changelog records a cache-hit-rate optimization on July 27 and a ZCode 3.6.5 entry dated August 3. Those entries document recent product activity, but they do not establish when Z.ai published the user milestone or reset the limits.
The X post also cited "98% cache" without defining the measure. Z.ai did not provide a workload, baseline or methodology, so the figure cannot be read as either a 98% cache-hit rate or a 98% reduction in compute. The July 27 changelog entry says ZCode optimized its cache-hit rate without disclosing a result.
Z.ai's Code Bench comparison reports a 0.01 percentage-point advantage for ZCode plus GLM-5.2 over Claude Code plus GLM-5.2 on full-task pass rate, alongside a displayed -0.00 percentage-point difference on checklist pass rate. The slide does not provide a sample size, test date or methodology.
The displayed differences amount to near-parity and do not establish an independent performance lead. That result could still help Z.ai if developers remain inside its product, where Z.ai can compete through subscription pricing, GLM integration, persistent state and remote execution.
ZCode turns Z.ai's model research into a developer product
Z.ai, previously known as Zhipu AI, was founded in 2019 as a Tsinghua University spinout and is based in Beijing. The founders brought research in knowledge engineering and language models into a commercial organization that now sells direct access to its models and coding tools.
Jie Tang, a Tsinghua professor, Z.ai co-founder and chief scientist, describes his research mission as teaching machines to think like humans. He has spent years working on the GLM model family, while his Tsinghua group has produced work spanning knowledge graphs, agent evaluation and large language models.
Li Juanzi, another Z.ai co-founder and Tsinghua professor, earned her doctorate in computer science from Tsinghua in 2000 after bachelor's and master's degrees at Shanxi University. Her research includes knowledge graphs, large models, question answering and dialogue generation.
Co-founder and CEO Zhang Peng earned a Tsinghua doctorate in 2018 and worked on GLM, the AMiner research platform and the XLORE knowledge graph before leading the commercialization of the research group's work.
Z.ai has financed that expansion with substantial outside capital. DigiTimes reported that Zhipu had raised more than CNY16 billion and that a roughly $400 million financing in 2024 included Prosperity7 Ventures. DigiTimes also associated Alibaba, Tencent, Xiaomi, Meituan, Ant Group and Chinese government-backed funds with Z.ai's financing. The supplied material does not establish Z.ai's current valuation.
ZCode competes with Anthropic's Claude Code, a terminal-based coding agent that can read repositories, edit files and run tests, and OpenAI's Codex, which supports parallel agents, background work and automation. Coder Agents offers a model-agnostic alternative that keeps planning, orchestration and execution in a customer's environment.
The quota reset lowers the immediate cost of testing longer workflows, while the self-reported 1 million-user milestone gives Z.ai a sizable top-of-funnel claim. Active use, completed tasks, retained subscriptions and inference cost will provide clearer measures of whether the founders have built a durable coding business around GLM.