GLM 5.2 in Claude Code -- Setup Tutorial

GLM 5.2
in Claude Code

Setup tutorial · 5x cheaper than Opus · Per-project switching · June 2026

glm-5.2claude-codevscodeopen-source

Provenance Labs · Gambit (Hermes)High Confidence

Executive Summary

GLM 5.2 is a 756B parameter open-source model you can route into Claude Code via Z.ai. 5x cheaper than Opus 4.8. Handles ~80% of knowledge work. One config file, per-project switching, works in VS Code extension.

5 Min Setup

Z.ai API key + one JSON config in .claude/settings.local.json. No CLI changes.

5x Cheaper

Input $1.40 vs $5. Output $4.40 vs $25. Plans from $16/mo.

Per-Project

Different config per folder. VS Code extension inherits automatically.

80/20 Rule

GLM for most tasks. Opus for heavy reasoning. Models are engines, harness is the car.

Setup: Steps 1-3

API key, config file, verify

Step 1: z.ai → create account → pick plan ($16/$64/$144) → API Keys → copy key.
Step 2: Drop config into .claude/settings.local.json.
Step 3: Open Claude Code → verify "GLM 5.2 1M context" in header.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "your-z-ai-key",
    "ANTHROPIC_API_KEY": "",
    "API_TIMEOUT_MS": "3000000",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.2",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.2",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.2",
    "ANTHROPIC_SMALL_FAST_MODEL": "glm-5.2",
    "CLAUDE_CODE_SUBAGENT_MODEL": "glm-5.2"
  }
}

All model defaults → glm-5.2. Timeout 50 min. Auth via Z.ai key.

Per-Project Switching + VS Code

One config per folder, zero friction

Claude Code is a harness. The model is the engine. Switch engines by changing which folder you open.

~/projects/
  glm-project/
    .claude/settings.local.json  ← GLM 5.2
  opus-project/
    (no config)                  ← Default Opus

Terminal: cd glm-project && claude → GLM. cd opus-project && claude → Opus.
VS Code: extension reads same .claude/settings.local.json per workspace. Open workspace in glm-project → uses GLM. Open in opus-project → default Opus.
No manual switching. The right model loads based on where you are.

When GLM 5.2 Wins

80% of your knowledge work, 5x cheaper

GLM 5.2: 3m 59s for a design task that took Opus 14m 59s. Equivalent quality. Creative, fast, cheap.

Creative one-shot builds: "Build me an interactive HTML page" → solid, often faster than Opus
Design work: landing pages nearly indistinguishable from Opus in side-by-side tests
Storm research: multi-agent research with subagents and verification -- skill orchestration matters more than model
Documentation + boilerplate: READMEs, configs, slide decks, reports
80% of daily tasks: anything that doesn't require deep multi-step reasoning

When Opus Is Still Necessary

The 10-20% that demands precision

GLM: 24 min vs. Opus: 5 min on complex reasoning. GLM missed duplicate record edge cases (true vs 1) that Opus caught.

Heavy reasoning: multi-step logic with deep context. GLM can be 5x slower on these.
Subtle edge cases: type coercion, null handling, boundary conditions
Complex debugging: intricate logic across multiple files
Strategic thinking: "Based on all this data, what matters and how do I apply it?"
Rule of thumb: start with GLM. If it struggles or takes >3x longer, switch to Opus.

Benchmarks

Where GLM 5.2 stacks up against closed-source models

GLM 5.2 is competitive with Opus 4.8, GPT 5.5, and Sonnet across major benchmarks. But Nate: "Take benchmarks with a grain of salt -- feel matters more."

Frontier SWE: GLM 5.2 beat GPT 5.5 on software engineering benchmark
Aider coding: beats Sonnet and Opus 4.7 on multiple evals
General knowledge: comparable to Opus 4.8 and GPT 5.5
Bottom line: GLM is a top-tier model by benchmarks but lacks Opus-level reasoning depth on complex multi-step problems. It's real, not a toy.

Pricing: The 5x Story

Why this matters for daily workflow economics

Model	Input/M tok	Output/M tok
Opus 4.8	$5.00	$25.00
GLM 5.2	$1.40	$4.40
Savings	3.6x	5.7x

Z.ai plans: $16/mo, $64/mo, $144/mo (yearly discount)
Nate's recommendation: $100 Claude plan + $64 Z.ai plan. Switch per task.
Heavy day of coding: ~5x cheaper with GLM for equivalent work

Risks & Gotchas

Quota limits: Z.ai plans have 5-hour and weekly quotas. Peak hours consume higher multiples.
Speed variance: GLM is unpredictable. Fast on design, slow on reasoning. Set API_TIMEOUT_MS: 3000000.
API compatibility: Z.ai emulates Anthropic's API. Edge cases may not be fully supported.
Subagent behavior: GLM subagents work but may differ from Opus. Test your workflows before committing.
Not self-hosted (yet): 756B params needs serious hardware. Most users rent from Z.ai -- still a dependency.
Fable lesson: closed-source models can be pulled overnight. Open-source = you own it. GLM 5.2 is a hedge.

Get Started

1
5-minute setup. z.ai account → API key → drop config into .claude/settings.local.json → test with /goal.
2
Set up per-project switching. glm-projects/ folder with GLM config. Main projects on default Opus. VS Code auto-detects.
3
Run a side-by-side. Take one real task. Run through GLM and Opus. Compare time, cost, quality. Build intuition.
4
Test with your skills. Verify custom skills, MCP servers, and agent teams work with GLM. Fix any breakage.
5
Track savings. Compare billing between GLM and Opus days. Target 3-5x cost reduction on daily knowledge work.

Sources

3 sources · 3 Tier 1

[1] GLM 5.2 in Claude Code -- Nate Herk, AI Automation (Tier 1)
[2] Z.ai -- GLM 5.2 API Console & Docs (Tier 1)
[3] Claude Code Settings -- Anthropic Docs (Tier 1)

Tier 1 = Primary source (tutorial, official docs, API platform) · Tier 2 = Community resource · Tier 3 = News

Generated June 21, 2026 · DeepSeek V4 Pro · Gambit / brief-to-slides v2.0