Z.ai Launches GLM-5.2 With a Usable 1M-Token Context, Two Thinking-Effort Levels, and No Benchmarks at Launch

GLM-5.2 is the latest release from Z.ai, marking the third iteration in the GLM-5 family. It arrives just months after GLM-5 (February 11), GLM-5-Turbo (March 15), and GLM-5.1 (April 7). This rapid cadence delivers four flagship-tier coding models in a single quarter.

What this means for developers and builders

The headline specification is a usable 1,000,000-token context window. Z.ai designates this variant as glm-5.2[1m]. Each response can generate up to 131,072 output tokens, representing roughly a fivefold increase over the 200,000-token window of GLM-5.1.

A 1M-token window fundamentally alters how an agent operates. It can retain an entire mid-sized repository in working memory—source code, tests, configuration, and conversation history—eliminating the constant summarisation forced by smaller limits.

The launch also introduces two distinct thinking-effort levels: High and Max. Z.ai advises using Max effort for complex, multi-step coding tasks. In Claude Code, the /effort command manages this setting, where options like xhigh, max, and ultracode all map to GLM-5.2’s Max effort mode.

Architecture and technical shifts

Z.ai did not disclose GLM-5.2’s specific architecture in its launch documentation. However, community analysis suggests the GLM-5 base is a 744-billion-parameter Mixture-of-Experts model, activating 40 billion parameters per token. GLM-5.1 retained this backbone with retargeted post-training.

Interactive Configuration Playground

The following visualiser demonstrates how to configure the model for specific agent workflows.

The missing benchmark data

A critical caveat: Z.ai published no benchmark scores for GLM-5.2 at launch. There are no figures for SWE-bench, Terminal-Bench, or the Code Arena. The announcement prioritised availability, context capabilities, and the open-source roadmap over performance metrics.

Specification Comparison: GLM-5.2 vs GLM-5.1

Practical use cases

• Whole-repository refactors: Load a mid-sized repository into a single context window. The agent tracks cross-file dependencies without needing to re-fetch data. Example: refactoring a 40-file Python data pipeline in one session.
• Long-horizon agent runs: GLM-5.2 targets sustained planning, execution, testing, and fixing loops. GLM-5.1 previously sustained roughly 1,700 agent steps in a single session, running autonomous loops for up to eight hours. GLM-5.2 inherits this trajectory, though specific numbers are pending.
• Drop-in Claude Code replacement: Swap only the base URL and model identifier. Keep your existing agent harness and workflow intact. This is vital when frontier API access is disrupted.
• Large-document analysis: Feed long specifications, logs, or transcripts exceeding 200K tokens. The 1M window accommodates material that smaller models would truncate.

Setup guide for GLM-5.2

For Claude Code, edit ~/.claude/settings.json. Point the Sonnet and Opus slots to the 1M variant. Raise the auto-compact window so the agent utilises 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, configure the endpoint via environment variables. The Anthropic-compatible endpoint accepts a base-URL swap.

Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.

Please enable JavaScript in your browser to complete this form.

Name *

First

Last

Name Email

Email *

Submit