Z.ai 发布 GLM-5.2，支持可用 1M Token 上下文窗口及两种思考强度

GLM-5.2’s standout spec is a 1,000,000-token context window. Z.ai labels the variant glm-5.2[1m] in its own configuration. Each response can return up to 131,072 output tokens. That is roughly a 5x jump from GLM-5.1’s 200,000-token window.

A 1M-token window changes how a coding agent works in practice. The agent can hold an entire mid-sized repository in working memory. That includes source files, tests, configuration, and conversation history. It avoids the constant summarization that smaller windows force.

The release also adds two thinking-effort levels: High and Max. Z.ai recommends Max effort for complex, multi-step coding work. In Claude Code, the /effort command controls this setting. The xhigh, max, and ultracode options all map to GLM-5.2’s Max effort.

Architecture and What Changed

Z.ai did not specify GLM-5.2’s architecture in its launch materials. But based on community notes, the GLM-5 base is a 744-billion-parameter Mixture-of-Experts model. It activates 40 billion parameters per token. GLM-5.1 kept that same backbone with retargeted post-training.

MTP Explainer Playground

Interactive Demo

GLM-5.2 Setup Generator & Context Visualizer

Pick your agent and effort mode. Copy the exact config. See what 1M tokens buys you.

1. Coding agent

2. Context window

3. Thinking effort

Your config

Context window: GLM-5.1 vs GLM-5.2

GLM-5.1

~200,000 tokens

GLM-5.2

1,000,000 tokens

GLM-5.2 at a glance

1,000,000

input tokens in one context window

max output tokens per response

5x

larger than GLM-5.1’s window

agentic tools supported day one

Config sourced from Z.ai developer docs · June 2026

The Benchmark Question

Here is the important caveat. Z.ai published no benchmark scores for GLM-5.2 at launch. There is no SWE-bench, Terminal-Bench, or Code Arena number yet. The announcement focused on availability, context, and the open-source roadmap.

Specification Comparison: GLM-5.2 vs GLM-5.1

Use Cases With Examples

• Whole-repository refactors: Load a mid-sized repo into one context window. The agent tracks cross-file dependencies without re-fetching. Example: refactor a 40-file Python data pipeline in a single session.
• Long-horizon agent runs: GLM-5.2 targets sustained plan, execute, test, fix loops. GLM-5.1 sustained roughly 1,700 agent steps in one session. It ran autonomous loops for up to eight hours. GLM-5.2 inherits that trajectory, though its own numbers are pending.
• Drop-in Claude Code replacement: Swap the base URL and model identifier only. Keep your existing agent harness and workflow. This matters when frontier API access is disrupted.
• Large-document analysis: Feed long specs, logs, or transcripts past 200K tokens. The 1M window holds material that smaller models truncate.

How to Set Up GLM-5.2

For Claude Code, edit ~/.claude/settings.json. Point the Sonnet and Opus slots at the 1M variant. Raise the auto-compact window so the agent uses the full context.

{
  "env": {
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-4.5-air",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.2[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.2[1m]"
  }
}

Alternatively, set the endpoint through environment variables. The Anthropic-compatible endpoint accepts a base-URL swap.

export ANTHROPIC_AUTH_TOKEN="your-zai-api-key"
export ANTHROPIC_BASE_URL="https://api.z.ai/api/anthropic"
export ANTHROPIC_DEFAULT_OPUS_MODEL="glm-5.2[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="glm-5.2[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-4.5-air"
claude

Then run /effort in a session and select max. Run /status to confirm GLM-5.2 is active. For Cline, choose the OpenAI Compatible provider. Set the base URL to https://api.z.ai/api/coding/paas/v4. Enter the custom model glm-5.2 and set context to 1,000,000.

GLM-5.2 is compatible with eight agentic coding tools from day one. The list includes Claude Code, Cline, OpenCode, and OpenClaw.

Key Takeaways

• Z.ai shipped GLM-5.2 on June 13, 2026, live immediately across all GLM Coding Plan tiers (Lite, Pro, Max, Team).
• 1M-token context window (glm-5.2[1m]) with up to 131,072 output tokens.
• No benchmarks were published at launch
• It drops into Claude Code, Cline, and OpenClaw via an Anthropic-compatible endpoint with just a base-URL and model swap.