How to Use GLM-5.2 in Claude Code and Cut Costs by Up to 50%

GLM-5.2 is an excellent coding model. Claude Code is an excellent coding-agent harness.

You can combine them.

This guide shows three practical ways to run GLM-5.2 inside Claude Code:

• OpenRouter, which is the best starting point if you want one routing layer and cost visibility.
• Z.AI direct, which is the cleanest provider-native setup.
• A local or internal proxy, which is the advanced setup for teams that need custom routing.

The cost case is straightforward. GLM-5.2 can be materially cheaper for coding-agent work, especially when the task does not require the most expensive default model. Optimize for cost per accepted coding outcome, not raw token volume.

Key Findings

Evidence and Methodology

OpenRouter lists the GLM-5.2 model ID as:

z-ai/glm-5.2

OpenRouter also lists GLM-5.2 with a 1M-token context window and pricing of $0.95 / $3 per 1M input/output tokens at the time of writing. That makes it a serious candidate for cost-sensitive coding-agent work.

List price is only the starting point. In a public GLM vs Opus cost thread, Sridhar Ramaswamy described a benchmark where GLM used about twice the tokens but still cost roughly half as much.

More tokens can still cost less.

For coding agents, measure the full session:

Concrete Operator Scenario

An engineering team already uses Claude Code every day. Engineers like the harness and want to keep it. Finance sees AI spend rising and asks whether there is a cheaper default for some classes of work.

The team keeps Claude Code and changes the model route.

They test GLM-5.2 for:

This is the operating model: keep Claude Code, route suitable work to GLM-5.2, and keep premium models for tasks where they earn their cost.

Measurement Approach

Run a small benchmark before changing defaults.

Use the same repository, prompt, acceptance criteria, and test commands across your current Claude Code model and GLM-5.2.

Track:

Treat lower model pricing as a hypothesis. The benchmark passes when cost per accepted task drops.

Path 1: OpenRouter

Use this path first if your team wants one API layer for model routing, usage visibility, and spend controls.

Create an OpenRouter API key, then run Claude Code with OpenRouter as the base URL:

export ANTHROPIC_BASE_URL="https://openrouter.ai/api"
export ANTHROPIC_AUTH_TOKEN="$OPENROUTER_API_KEY"
export ANTHROPIC_MODEL="z-ai/glm-5.2"
claude

For persistent setup, add it to ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://openrouter.ai/api",
    "ANTHROPIC_AUTH_TOKEN": "YOUR_OPENROUTER_API_KEY",
    "ANTHROPIC_MODEL": "z-ai/glm-5.2"
  }
}

Then restart Claude Code and run:

/status

Confirm that Claude Code is using the expected route. Then run a real coding task.

OpenRouter caveat: its Claude Code documentation says Claude Code support through OpenRouter is guaranteed only with Anthropic first-party providers. GLM-5.2 may work through OpenRouter, but you should treat it as a compatibility test until your own file-editing, bash, test, and tool-calling checks pass.

Use OpenRouter when you want:

• centralized cost reporting,
• one key across multiple providers,
• budget controls,
• model comparison,
• routing and fallback.

Use it for team-wide Claude Code defaults only after tool calling is verified.

Path 2: Z.AI Direct

Use this path if you want the cleanest GLM-native setup.

Z.AI documents an Anthropic-compatible Claude Code endpoint:

https://api.z.ai/api/anthropic

Add this to ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "YOUR_ZAI_API_KEY",
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "API_TIMEOUT_MS": "3000000",
    "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]"
  }
}

Restart Claude Code:

claude

Check:

/status

This is the safest GLM-specific path because it uses an Anthropic-compatible endpoint instead of forcing Claude Code through an OpenAI-compatible API.

Path 3: Local or Internal Proxy

Use this path if your team already runs an AI gateway or needs custom routing.

Claude Code expects Anthropic-style behavior. Some GLM endpoints and gateways expose OpenAI-compatible Chat Completions instead. Put a proxy between Claude Code and an OpenAI-compatible endpoint so the request and tool-call formats are translated correctly.

The shape is:

Claude Code
  -> Anthropic-compatible proxy
  -> OpenAI-compatible GLM endpoint
  -> GLM-5.2

Use this setup when you need:

• model aliases,
• team-level attribution,
• internal logging,
• policy controls,
• caching,
• retries,
• fallback routing.

This is the most flexible option and the easiest one to break. Test it carefully.

Caveats And Failure Modes

Most failures come from the adapter layer.

The practical rule: if search, edit, bash, git, tests, and long streaming responses work, the setup is worth benchmarking. If any of those fail, fix the route before judging GLM-5.2.

What To Do Next

Start with OpenRouter if you care about routing and visibility. Start with Z.AI direct if you want the cleanest GLM-specific setup.

Run this five-task test:

• Ask Claude Code to inspect a real repo and explain the relevant files.
• Ask it to make a small implementation change.
• Ask it to update or add tests.
• Ask it to run tests and fix failures.
• Ask it to summarize the final diff and risks.

Compare GLM-5.2 against your current default model on:

• total session cost,
• retry count,
• latency,
• tool-call success,
• test success,
• review rewrite,
• accepted outcome rate.

If GLM-5.2 cuts cost without increasing retries or review burden, make it the default for those task classes. Keep stronger models for work that needs them.

Related Pages

• What Is AI-Native Developer Intelligence?
• What Is Engineer-Agent Effectiveness?
• What Is Agent Readiness?
• What Is Prompt Fluency?
• What Is Token Cost Effectiveness for AI Coding?

FAQ

Can I use GLM-5.2 inside Claude Code?

Yes. Use OpenRouter, Z.AI's Anthropic-compatible endpoint, or a proxy that translates Claude Code's Anthropic-style requests.

Which setup should I try first?

Try OpenRouter first if you want routing and cost visibility. Try Z.AI direct first if you want the simplest GLM-specific route.

What is the OpenRouter model ID?

Use z-ai/glm-5.2.

What is the biggest caveat?

Tool calling. Verify file search, edits, bash commands, tests, git, and long streaming responses before using the setup for serious work.

Can this really cut costs by up to 50%?

Yes, for the right task classes. The savings depend on model pricing, retries, context size, cache behavior, and review quality. Measure cost per accepted coding task.