On August 11, Zhipu AI announced a major upgrade to ZCode, its homegrown coding harness, officially launching four new features: Goal, Subagents, Remote Control, and off-peak tasks. As a thank-you for passing the 1 million user milestone, all GLM Coding Plan quotas were reset to full at 13:00 that day. Original link: https://zcode.z.ai/cn
Not the Same Species as Claude Code
ZCode is Zhipu's self-developed domestic coding harness — not an IDE wrapper, not a ChatGPT wrapper, but a framework of "context management + tool calling + task scheduling + caching + result verification" deeply optimized for GLM models. Zhipu internally defines the harness as the middle layer that "determines how much of a model's capability can actually be realized" — the same GLM-5.2 run inside ZCode and Claude Code shows significantly different real-world performance.
Zhipu built Z.ai Code Bench to quantify this. The benchmark is based on real user requirement scenarios across Full Stack, Bug Fix, and Feature Implementation subtasks, constructing simulated local programming environments evaluated on four dimensions: regression testing, new feature testing, frontend simulated interaction, and code quality assessment. Its stated design goal is to "avoid the data contamination that public leaderboards may suffer" — a direct response to the gamification of SWE-bench-style evaluations.
The Real Gap: 2.39%
On Z.ai Code Bench, GLM-5.2 + ZCode scores 2.39% higher than GLM-5.2 + Claude Code on overall task pass rate — but on checklist pass rate, ZCode is 1.22% lower.
This combination is interesting: ZCode is stronger at "getting the task done end-to-end," while Claude Code is stronger at "strictly matching the checklist at every step." Zhipu's own interpretation: ZCode's advantage concentrates on complex tasks requiring cross-file, cross-stage collaboration with final acceptance. For real production scenarios (a PR touching 5 files + passing CI + unit tests), ZCode's end-to-end pass rate is higher — but it is looser about per-step strictness such as log formats or naming conventions.
The Economics of a 98% Cache Hit Rate
ZCode applies targeted optimization to context cache reuse, with a cache hit rate above 98% (official figure: 98.10%). This raises GLM Coding Plan's effective token volume by about 30% — repeated contexts are billed at a lower credit coefficient, so developers run more tasks for the same money.
Combined with the limited-time 1.5x quota bonus (through August 31), total usable volume approaches 1.8x regular quotas. It's a price signal in the coding subscription race: running GLM in ZCode before end of August is nearly half price.
Goal Mode: From Babysitting Agents to Setting Targets
The core of this upgrade. The old coding agent workflow:
> Developer sends request → wait for reply → push forward → tests fail → ask again → discover omissions → add context → push again
Goal mode becomes:
> Set "first-screen load ≤ 2s" + "all existing tests pass" → ZCode decomposes automatically → edits code → runs commands → executes tests → checks results → next round if not met → until done
Each round's progress, duration, and results are visualized in the Goal panel. It imports the "sparse reward + many trials" RL training paradigm into the product form of a coding agent; code review shifts to after goal completion.
Subagents: A Parallel Development Team
Under Goal mode, Subagents turn one agent into a parallel team. Two built-in types:
- General-purpose: modify code, fix issues, run commands
- Explore: read-only exploration — locate code, analyze call chains, gather pre-change evidence
- Entry: WeChat, Feishu, Lark — scan a QR code or open a browser link
- Phone can: view progress, send instructions, create new tasks, reconnect to a disconnected workspace
- Phone cannot: upload code or commands — they keep running in the desktop's original environment (local / SSH / WSL / Docker)
- Key point: no cloud environment is created, and projects are never synced to the phone
- Overall task pass rate, GLM-5.2 + ZCode vs + Claude Code: +2.39%
- Checklist pass rate: -1.22%
- Cache hit rate: 98.10%
- Effective token volume gain: ~30%
- Limited-time quota bonus: 1.5x (through 2026-08-31)
- Combined effective usage: ~1.8x regular quota
- ZCode users: 1M+ (announced 2026-08-11)
- Built-in subagent types: General-purpose, Explore
- Remote Control entries: WeChat, Feishu, Lark, browser link
- Workspace compatibility: local / SSH / WSL / Docker
- Foundation model: GLM-5.2 (Zhipu in-house)
- Harness: ZCode (Zhipu in-house, optimized for GLM)
- Subscription: 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 (Zhipu official WeChat account)
- https://finance.sina.com.cn/tech/digi/2026-08-11/doc-inimxmep7594993.shtml (Sina Finance, via ITHome)
- https://www.chinaz.com/ainews/30242.shtml (Chinaz)
- https://www.163.com/dy/article/L42IHT8M053469RG.html (NetEase Tech)
Users can define custom subagents with configured model, permission scope, and prompts, invoked directly in chat via "/". In cross-file refactoring, complex bug hunting, and full-stack feature work, "one agent does everything" is replaced by "multiple agents divide the labor" — reducing long-task pressure on the main conversation context.
Remote Control: Phone Access Without Sending Code
Remote Control temporarily exposes the current ZCode desktop window to a phone:
It's the "remote work" version of a coding agent — push your agent forward away from your desk while the code never leaves your machine.
Off-Peak Tasks: Long Jobs at Off-Peak Hours, No Credits Charged
Off-peak tasks let users submit non-urgent, long-running jobs (massive test runs, code optimization) that ZCode executes automatically during low-peak hours, without deducting GLM Coding Plan credits within entitlement limits. It brings cloud computing's "off-peak discount" to the product layer — the marginal cost of long tasks approaches zero.
Key Numbers
Zhipu's "Foundation Model + Harness + Subscription" Trinity
ZCode is not another Claude Code clone. It closes the loop of a domestic foundation model + domestic harness + domestic subscription:
Zhipu has been filling in its base capabilities — models, agents, MaaS, industry ecosystem — in parallel. ZCode is the entry point closest to developers: a million-user base, cache economics, and Goal/Subagents autonomous delivery.
Compared to Claude Code: Anthropic's harness, Anthropic's model, Anthropic's subscription — three layers coupled in one company. ZCode brings all three in-house — but must face the moat Claude Code already holds in Western developers' minds.
The most interesting signal to watch: whether Zhipu will open up Z.ai Code Bench (currently internal) so third parties can independently verify the 2.39% gap of GLM-5.2 + ZCode vs + Claude Code. If it goes public and third parties reproduce it, it becomes a turning point for domestic AI coding tools; if it stays internal, "2.39% ahead of Claude Code" should be taken with a grain of salt.
---
Sources