Overview
On August 11, Zhipu AI announced a major upgrade to ZCode, its proprietary Coding Harness built around the GLM-5.2 model, marking a one-million-user milestone. The update introduces four flagship capabilities — Goal mode, Subagents, Remote Control, and Idle Tasks — and resets all GLM Coding Plan user quotas to full at 13:00 the same day.
Source: https://zcode.z.ai/cn
What Makes ZCode Different from Claude Code
ZCode is not an IDE wrapper or a ChatGPT wrapper. It is a domestic Coding Harness purpose-built for the GLM family, integrating context management, tool calling, task scheduling, caching, and result validation. Zhipu defines a Harness as the middle layer that determines how much of a model's latent capability is actually realized — the same GLM-5.2 produces notably different outcomes when routed through ZCode versus Claude Code.
For measurement, Zhipu built Z.ai Code Bench, an internal benchmark derived from real user scenarios across Full Stack, Bug Fix, and Feature Implementation subtasks. It runs against simulated local programming environments and scores submissions along four axes: regression testing, new-feature testing, frontend simulation interaction, and code quality. The stated goal is to avoid data contamination that affects public leaderboards such as SWE-bench.
The 2.39% Real-World Gap
On Z.ai Code Bench, GLM-5.2 + ZCode outperforms GLM-5.2 + Claude Code by +2.39% on overall task pass rate, but trails by -1.22% on checklist pass rate.
The contrast is instructive. ZCode is stronger at end-to-end task completion across multiple files and verification stages; Claude Code is stricter at matching every checklist step exactly. For real production work — a PR that touches five files, runs CI, and passes unit tests — ZCode's higher end-to-end pass rate is meaningful, while Claude Code enforces tighter format and naming compliance.
Cache Hit Rate and the Economics of 98%
ZCode dedicates significant engineering to context-cache reuse. The official figure is 98.10% cache hit rate, which lifts the effective token count available through GLM Coding Plan by roughly 30% because repeated context is billed at a lower credit coefficient. Combined with a 1.5x limited-time credit bonus valid through August 31, 2026, total usable capacity approaches ~1.8x normal quota.
Goal Mode: Set a Target and Walk Away
Traditional coding-agent workflow:
> Developer issues request → waits for reply → nudges agent → test fails → follow-up → finds omission → adds context → re-prompts
Goal mode replaces this with:
> Define targets (e.g., first-paint ≤ 2s, all existing tests pass) → ZCode decomposes the task → edits code → runs commands → executes tests → inspects results → loops if criteria unmet → completes
Per-round progress, timing, and execution results are visualized in the Goal panel. The paradigm mirrors reinforcement learning — sparse reward, many trials — applied to product-level coding agents. Code review is deferred until after Goal completion.
Subagents: A Parallel Development Team
Within Goal mode, Subagents transform a single agent into a coordinated team. Two built-in variants:
- General-purpose: edits code, fixes issues, runs commands
- Explore: read-only investigation — locates code, traces call chains, gathers pre-change evidence
- Entry points: WeChat, Feishu, Lark, or browser link via QR code
- Mobile can: view progress, send instructions, create new tasks, reconnect a dropped workspace
- Mobile cannot: code and commands never leave the desktop environment (local / SSH / WSL / Docker)
- Key property: no cloud environment is created and no project is synced to the phone
- End-to-end task pass rate, GLM-5.2 + ZCode vs + Claude Code: +2.39%
- Checklist pass rate: -1.22%
- Cache hit rate: 98.10%
- Effective token capacity lift: ~30%
- Limited-time credit bonus: 1.5x (through 2026-08-31)
- Combined effective usage: ~1.8x normal quota
- ZCode user count: 1,000,000+ (announced 2026-08-11)
- Built-in subagent types: General-purpose, Explore
- Remote Control entry points: WeChat, Feishu, Lark, browser link
- Workspace compatibility: local / SSH / WSL / Docker
- Foundation model: GLM-5.2 (Zhipu in-house)
- Harness: ZCode (Zhipu in-house, GLM-optimized)
- Subscription service: GLM Coding Plan (Zhipu-operated)
- https://zcode.z.ai/cn
- https://mp.weixin.qq.com/s?__biz=MzkyMzI3NzQ0Mg%3D%3D&mid=2247494052&idx=1&sn=ee3ab3d0f4550e9120927c53a27522c9
- https://finance.sina.com.cn/tech/digi/2026-08-11/doc-inimxmep7594993.shtml
- https://www.chinaz.com/ainews/30242.shtml
- https://www.163.com/dy/article/L42IHT8M053469RG.html
Users can define custom subagents with configurable models, permission scope, and prompts, then invoke them in chat via "/". For cross-file refactors, complex bug hunts, and full-stack feature delivery, parallel subagents reduce the long-task burden on the main conversation context.
Remote Control: Mobile Oversight Without Code Upload
Remote Control temporarily exposes the active desktop ZCode session to a mobile device:
This is "remote work" for coding agents — operators can keep pushing the agent from anywhere while code stays on the workstation.
Idle Tasks: Run Long Jobs Off-Peak for Free
Idle Tasks lets users submit non-urgent, long-running work — large test suites, code optimization — to be scheduled during low-traffic windows. Execution does not deduct GLM Coding Plan credits within plan entitlements, effectively applying cloud-spot-style discounts to coding-agent workloads.
Key Numbers
Zhipu's Foundation Model + Harness + Subscription Stack
ZCode is positioned as a fully integrated domestic loop:
By contrast, Claude Code couples Anthropic's harness, Anthropic's model, and Anthropic's subscription inside a single vendor. ZCode decomposes the same stack across Zhipu's own layers, but must compete against Claude Code's entrenched mindshare among Western developers.
The signal most worth watching: whether Zhipu opens Z.ai Code Bench to independent third-party verification. If the +2.39% lead over Claude Code can be reproduced externally, it would mark a turning point for domestic AI coding tools. If the benchmark stays internal, the claim will be discounted accordingly.