ZCode 3.0: Z.ai's Open-Weight Answer to Claude Code and Codex CLI

In mid-2026, the AI-assisted development market gained a serious new contender. Z.ai (the international brand of China's Zhipu AI) released ZCode 3.0 on June 13, 2026 — not just another IDE plugin, but a full Agentic Development Environment (ADE) designed around its open-weight flagship model, GLM-5.2 .

ZCode is free to download. What you pay for is access to the GLM-5.2 model through a subscription called the GLM Coding Plan. The result is a development experience that directly challenges Anthropic's Claude Code, OpenAI's Codex CLI, and startup darling Cursor, at a fraction of their cost .

What is ZCode? More Than an IDE

ZCode is not a traditional IDE where you manually write code. It is an ADE — the agent is the primary actor. You describe your intent in natural language, and a proprietary ZCode Agent handles file editing, terminal commands, Git operations, and even runs a live browser preview .

Unlike its earlier versions (which were lightweight wrappers around third-party agents like Claude Code), ZCode 3.0 switched to a fully self-developed ZCode Agent core, optimized specifically for long-horizon reasoning and tool calling within the GLM ecosystem . The Chinese tech press described it as Z.ai's "bombshell" release, signaling a complete pivot to a proprietary stack .

Key Features

• Full-context awareness: The agent understands the entire project structure — files, directories, dependencies, and UI elements — without you needing to manually pass parameters .
• Multi-agent collaboration: You can spawn sub-agents for parallel work, create reusable "Skills" (saved agent workflows), and manage complex tasks in a multi-agent workspace .
• Model Context Protocol (MCP) integration: Connect external tools and services for web search, file reading, or vision analysis .
• BYOK (Bring Your Own Key): You are not locked into GLM. ZCode supports Anthropic Claude, OpenAI models, and any compatible endpoint, making it a universal harness .
• SSH remote development: Work on code running on remote servers directly from the desktop app .
• Mobile and chat control: Send tasks and monitor progress via WeChat or Telegram .
• New in 3.0: A token usage dashboard, a Git branch graph, and Zread — a built-in project documentation reader .
• Platforms: Native desktop apps for macOS and Windows, with Linux in beta .

The Engine: GLM-5.2 Model Specs and Performance

ZCode's power comes from GLM-5.2, Z.ai's coding-first, agent-optimized large language model launched on the same day as ZCode 3.0 . It was later released under a permissive MIT open-source license, making it the only frontier-class coding model with open weights .

Model Specs

• Parameters: 750 billion total, with approximately 40 billion activated per token via a Mixture-of-Experts (MoE) architecture .
• Context window: A usable 1 million tokens, supporting whole-codebase reasoning .
• Thinking modes: Two modes — High for most tasks and a deeper Max mode for complex builds .
• API Pricing: Roughly $2 per million input tokens and $4–6 per million output tokens .

Benchmark Performance

The headline finding is that GLM-5.2 matches or beats GPT-5.5 on several key coding evals and comes within a hair of Claude Opus 4.8 on two specific agentic benchmarks .

Benchmark	GLM-5.2	Claude Opus 4.8	GPT-5.5
SWE-bench Pro	62.1%	69.2%	58.6%
FrontierSWE	74.4%	75.1%	~72.6%
Terminal-Bench 2.1	81.0%	85.0%	—
MCP-Atlas	77.0%	77.8%	—
SWE-bench Verified	~70%	88.6%	—
IDOR Detection (Semgrep)	39% F1	32% F1 (Claude Code)	—

Sources:

Key takeaways from the benchmarks:

• GLM-5.2 scores 62.1% on SWE-bench Pro — beating GPT-5.5 (58.6%) but trailing Claude Opus 4.8 (69.2%) .
• On FrontierSWE (a test of long-horizon, multi-step coding), the gap to Opus narrows to just 0.7 points (75.1% vs 74.4%) .
• On MCP-Atlas, the margin is 0.8 points (77.8% vs 77.0%) .
• In an independent cybersecurity benchmark by Semgrep, GLM-5.2 outperformed Claude Code on IDOR detection, scoring 39% F1 versus 32% F1, at roughly $0.17 per vulnerability found .

Important caveat from independent testers: While GLM-5.2 is strong on architecture review, correctness, and broad system-level problems, Claude Opus 4.8 still wins on "deeper exploit-style bug hunting, front-end taste, and the kind of polish that matters when the output is going live" . One developer who swapped Opus for GLM-5.2 inside their Claude Code workflow called it "genuinely useful" but noted Opus remains ahead for the hardest agentic tasks .

Pricing: The GLM Coding Plan

The ZCode desktop application is free. To use any AI agent features — including GLM-5.2 — you need a GLM Coding Plan subscription. This is a flat monthly fee with a prompt-based quota that refreshes every 5 hours .

Tier	Base Monthly Price	Typical Quota (5-hour / Weekly)	Best For
Lite	~$18	~80 prompts / ~400 prompts	Light daily use, exploration
Pro	~$72	~400 prompts / ~2,000 prompts	Full-time individual developers
Max	~$160	~1,600 prompts / ~8,000 prompts	Heavy agentic and long-context work
Team	Custom	Custom	Organizations

Sources:

Pricing notes:

• Discounts: Z.ai offers roughly 10% off for monthly billing, 20% for quarterly, and 30% for annual commitments. On the annual plan, Lite drops to ~$12.60/month, Pro to ~$50.40, and Max to ~$112 .
• Promotions: Through July 31, 2026, a promotion meters GLM-5.2 usage at a 0.67 factor, effectively giving subscribers ~1.5x usable quota .
• No lock-in: ZCode also supports BYOK. You can use Anthropic or OpenAI API keys directly, bypassing the Coding Plan if you prefer .

Competitive Positioning in the 2026 Market

ZCode + GLM-5.2 enters a market dominated by Anthropic's Claude Code, OpenAI's Codex CLI, and Cursor. Its strategy rests on four pillars:

1. Open-Weight Disruption

GLM-5.2 is MIT-licensed. Unlike the closed models powering Codex and Claude Code, developers can self-host, fine-tune, and inspect the model . This appeals to organizations with strict data sovereignty requirements and to the open-source community.

2. Aggressive Pricing

The GLM Coding Plan starts at $18/month, undercutting Anthropic's $200/month Claude Max plan by an order of magnitude . Business Insider called ZCode "inexpensive" compared to competitors . For context, the API token pricing of GLM-5.2 ($2/$4–6 per million tokens) is 3.6x to 5.7x cheaper than Claude Opus 4.8 .

3. Geopolitical Positioning

Z.ai explicitly competes with Anthropic amid the latter's restrictions on Chinese users. The South China Morning Post reported that Z.ai is "heating up its rivalry with US firm Anthropic" after Claude Code removed code that tracked Chinese users . ZCode provides a domestic alternative for the Chinese market while going global under the Z.ai brand.

4. Flexible Ecosystem (BYOK + MCP)

Unlike Cursor or Claude Code, ZCode is not a walled garden. You can switch models within the same environment, connect MCP servers, and even control it from a phone . This makes it an attractive "harness" for developers who want to experiment with different models without switching tools.

The Bottom Line

ZCode 3.0 is the most credible open-weight alternative to the Claude Code / Codex workflow ecosystem. Its strengths are open licensing, aggressive pricing, a flexible multi-model approach, and a self-developed agent core tuned for long tasks. Its weakness is that Claude Opus 4.8 remains the performance leader on the hardest coding benchmarks by a significant margin .

Z.ai's next model, GLM-5.5, is expected in August 2026 and could further close the gap . For developers who prioritize cost, model flexibility, and open-source control over peak benchmark performance, ZCode is already a compelling choice.