ZCode Launches Free - GLM-5.2 Comes for Claude Code
Z.ai's free agentic IDE ships Goal Mode, multi-agent coordination, and pricing up to 82% cheaper than Claude Code - with a concrete China data law risk that teams need to weigh.
6 min read
Seven months ago, the real choices for an agentic coding environment were Claude Code, Cursor, and GitHub Copilot. On July 2, Beijing-based Z.ai added a fourth: ZCode, a free desktop IDE built on GLM-5.2 that currently sits at #2 on Code Arena globally and undercuts Claude Code's API pricing by up to 82%. There's a catch, and it isn't small.
Key Specs
| Spec | Value |
|---|---|
| Model | GLM-5.2 (744B MoE, 40B active) |
| Context window | 1M tokens |
| License | MIT (open weights) |
| Release | July 2, 2026 (v3.3.0 current) |
| Platforms | macOS, Windows, Linux |
| Free trial | 5M tokens/day for 5 days |
The Goal Architecture
ZCode doesn't present itself as an editor with an AI sidebar. It's an agent-first environment where the agent loop is the primary interface - the file tree is secondary.
Goal Mode
The central abstraction is /goal. Type /goal refactor authentication module to use JWT and ZCode plans the work, writes the code, runs tests, and keeps iterating until the goal is met or a stopping condition triggers. A status panel tracks elapsed time, token consumption, and iteration count throughout. You can pause, redirect, or clear a goal without resetting the session.
That continuity matters for large codebases. The 1M-token context window means ZCode doesn't lose track of a 200-file repository halfway through a refactor - a real failure mode for tools capped at 128K that start hallucinating module names after a few hours of work.
Five Autonomy Modes
ZCode ships five execution modes, cycled with Shift+Tab:
- Default Mode - balanced confirmation prompts
- Confirm Before Changes - requires approval before any file edit or command
- Auto Edit - file changes happen automatically, commands need manual approval
- Plan Mode - shows the full plan before implementation starts
- Full Access - minimal interruptions, highest autonomy
After the JadePuffer incident, where an autonomous agent chained exploitation steps without per-step human review, explicit autonomy controls feel less like an UI nicety and more like a safety requirement. ZCode at least gives teams the dial.
Custom subagents landed in v3.2.0 on June 29. They're defined as plain Markdown files at ~/.zcode/agents/:
---
name: test-runner
model: glm-5.2
tools: [terminal, file_browser]
---
You run test suites. Given a file or directory, identify relevant tests,
run them, and return a summary of failures with line numbers.
The agent is then callable with /agent test-runner. Place the file in .zcode/agents/ at the repo root to scope it to a project. The project also supports an optional AGENTS.md file for session-level conventions, similar in concept to Claude Code's CLAUDE.md.
What GLM-5.2 Brings
The underlying model is GLM-5.2, a 744B mixture-of-experts model with 40B parameters active per token, trained on Huawei silicon. We covered the model's open-weight launch in June. Benchmark positions relative to its primary competitors:
| Benchmark | GLM-5.2 | Claude Opus 4.8 |
|---|---|---|
| SWE-Bench Pro | 62.1% | 69.2% |
| FrontierSWE | 74.4% | 75.1% |
| Code Arena rank | #2 | not ranked |
The 7-point gap on SWE-Bench Pro is real. The FrontierSWE gap is 0.7 points. The two benchmarks measure different things - FrontierSWE highlights longer-horizon agentic tasks, which is where GLM-5.2's 1M context and ZCode's Goal architecture are designed to compete. GPT-5.5 scores 72.6% on FrontierSWE, putting GLM-5.2 ahead on that specific benchmark.
ZCode targets developers running complex, multi-hour agentic tasks where context retention matters as much as raw benchmark scores.
Source: pexels.com
Self-Hosting Reality
GLM-5.2 is MIT-licensed, and Z.ai released the weights on Hugging Face on June 17 - the same week US export controls pulled Fable 5 and Mythos 5 offline. The timing was intentional. Z.ai's public messaging framed the MIT license as a hedge against vendor dependency, specifically for teams that needed AI tooling not subject to US export control changes.
The math on self-hosting is harder. Running GLM-5.2 at full precision requires approximately 1.5TB of GPU memory. That's around 19 x H100 80GB cards at minimum - infrastructure most engineering teams don't have and can't justify for a coding assistant.
Pricing Breakdown
| Plan | Monthly cost |
|---|---|
| ZCode Lite | $16.20 |
| ZCode Pro | $64.80 |
| ZCode Max | $144 |
| Claude Code Max | $200 |
| Cursor Pro | $20 |
The API rate for teams using GLM-5.2 directly: $1.40 per million input tokens, $4.40 per million output tokens. Claude Opus 4.8 runs $5/$25 per million tokens, making GLM-5.2 72% cheaper on input and 82% cheaper on output. For teams running agentic coding loops at volume, that spread adds up quickly.
ZCode also ships BYOK (bring-your-own-key) support for Claude, Gemini, and OpenAI models. This is covered separately in the next section.
Platform availability: macOS (Apple Silicon and Intel), Windows (64-bit and ARM64), and Linux (x64 and ARM64 in beta, available as .deb and AppImage). The project shipped eight releases between June 26 and July 3, suggesting active development velocity.
Where It Falls Short
Every API call to Z.ai's cloud routes through Beijing-incorporated infrastructure subject to China's National Intelligence Law.
Source: pexels.com
The China Data Law
Every ZCode session that uses Z.ai's API routes through infrastructure operated by a Beijing-incorporated company. China's National Intelligence Law, enacted in 2017, requires all Chinese organizations to cooperate with state intelligence requests on demand. Article 7 states that companies must "support, assist, and cooperate with state intelligence work in accordance with law." The US Department of Homeland Security has specifically warned that this framework can compel Chinese companies to disclose data from foreign users, including code and prompts.
For non-sensitive personal projects, this is a calculated risk teams can assess for themselves. For code involving customer data, proprietary business logic, or security systems, routing prompts through Z.ai's API is a different calculation.
The BYOK Caveat
ZCode supports BYOK for Claude, Gemini, and OpenAI models. This might look like a workaround for the data law concern. It isn't. Even when using a non-GLM model via BYOK, your codebase, file structure, terminal output, and Git history are still processed by ZCode's orchestration layer on Z.ai's servers. BYOK changes which API generates the text. It doesn't change which server sees your code.
The Self-Hosting Gap
The sovereignty argument for MIT-licensed weights only holds if teams can actually run them. At 1.5TB VRAM, self-hosting is out of reach for most engineering organizations. Teams with that infrastructure also don't need ZCode's pricing advantages. The gap between "MIT-licensed open weights" and "practically self-hostable" is significant here.
ZCode is the most capable open-weight coding agent to ship in 2026, and the pricing pressure it puts on Cursor and Claude Code is real. The China data law concern isn't abstract - it's a legal obligation that applies to every API call, no matter which model you select via BYOK. Teams evaluating ZCode should make that trade-off explicitly. For a full breakdown of the underlying model, see the GLM-5.2 model profile. For a comparison of Western alternatives, see our Claude Code vs Cursor vs Codex breakdown.
Sources:
