Run GLM-5.2 Inside Claude Code: The Full Setup Guide

To run GLM-5.2 inside Claude Code, you need exactly two environment variables — an auth token and a base URL — plus one detail most setup guides get wrong: the documented model ID is glm-5.2[1m], brackets included, not bare glm-5.2. Drop the suffix and everything still appears to work, but your sessions quietly run against a smaller default context instead of the model’s full 1M-token window.

The reason this setup is worth thirty minutes of your time is the economics. GLM-5.2, announced June 13, 2026 with MIT-licensed open weights following on June 16, lands near-frontier on many single-shot coding benchmarks at a fraction of Claude’s API cost — while still trailing Claude Opus 4.8 on sustained long-horizon agent work, by the vendor’s own published numbers. That combination makes it a strong second engine inside the harness you already use, not a wholesale replacement.

This guide covers how the redirect actually works, all three official setup paths, the exact config block from Z.ai’s docs, the troubleshooting steps that fix nearly every failed install, an evidence-based task-routing table, and the honest cost math — by slot, not just the flattering Opus comparison.

Claude Code reads a handful of environment variables at launch. Set ANTHROPIC_BASE_URL to https://api.z.ai/api/anthropic and every request the harness makes — plans, edits, tool calls — routes to Z.ai’s Anthropic-compatible endpoint instead of Anthropic’s API. ANTHROPIC_AUTH_TOKEN carries your Z.ai key. Three more variables remap the model slots Claude Code exposes: the Opus and Sonnet slots to glm-5.2[1m], and the lightweight Haiku slot to glm-4.7, per the official Claude Code setup page.

Be clear about what this is: a whole-session swap, not per-prompt routing. Once the base URL points at Z.ai, there is no native way to send one prompt to GLM-5.2 and the next to a Claude model inside the same running session. The pattern several independent guides converge on is per-task instead — keep your native Claude setup untouched as the default, and add a small shell function that launches a GLM-backed session on demand:

# ~/.zshrc — native Claude stays default; launch GLM on demand
glm() {
  ANTHROPIC_AUTH_TOKEN="your-zai-api-key" \
  ANTHROPIC_BASE_URL="https://api.z.ai/api/anthropic" \
  ANTHROPIC_DEFAULT_OPUS_MODEL="glm-5.2[1m]" \
  ANTHROPIC_DEFAULT_SONNET_MODEL="glm-5.2[1m]" \
  ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-4.7" \
  claude "$@"
}

One more distinction before you generate a key. A GLM Coding Plan key bills as a flat monthly fee with prompt-based quota and is restricted to officially supported coding tools. A standard pay-as-you-go key works with any application and bills per token at list rates. Both plug into the same ANTHROPIC_AUTH_TOKEN slot — the difference is which kind of key you generate on Z.ai’s side, not a different Claude Code config. If you plan to run GLM-5.2 as an executor under a Claude orchestrator, our guide to pairing Fable 5 as orchestrator with GLM-5.2 as executor builds on exactly this per-task launch pattern.

Most walkthroughs show exactly one method. Z.ai documents three, and they suit genuinely different situations — this is a decision, not a list of synonyms.

Paths A and B in full:

# Path A — interactive wizard
npx @z_ai/coding-helper

# Path B — automated script (macOS / Linux only)
curl -O "https://cdn.bigmodel.cn/install/claude_code_zai_env.sh"
bash ./claude_code_zai_env.sh

The wizard’s npm package describes itself as “A CLI helper for GLM Coding Plan Users to manage coding tools like claude-code.” It sits at v0.0.7 across 37 published versions and predates GLM-5.2 itself — it is a general-purpose Coding Plan manager, not a model-specific installer. Beyond the wizard, the CLI carries a few commands worth knowing: chelper auth for interactive key entry, chelper auth reload claude to push updated plan info into your Claude Code config without re-running the whole flow, chelper auth revoke to delete a saved key, and chelper doctor for a health check.

This is the manual-configuration block as documented on Z.ai’s Claude Code page, retrieved July 3, 2026. Copy the model strings exactly — quoted, brackets included:

// ~/.claude/settings.json
{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "your-zai-api-key",
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.2[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.2[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-4.7"
  }
}

The same docs page recommends three optional variables. The long timeout matters because GLM’s serving can be slower at peak; the auto-compact window should match the 1M context so Claude Code’s compaction doesn’t trigger prematurely; the third trims telemetry and other nonessential requests:

// Optional — recommended on the same docs page
"API_TIMEOUT_MS": "3000000",
"CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000",
"CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"

Whichever path you took, the next step is the same and it is not optional: close every Claude Code window and open a new terminal. The docs state plainly that a new window is required for the changes to take effect — environment variables load at launch, so an already-open session never sees your edits. Then run claude. On first contact with the non-Anthropic base URL, Claude Code asks whether you want to use this API key — select Yes, exactly as the docs instruct.

Version hygiene matters more than it looks. Z.ai’s docs state: “We have verified compatibility with Claude Code 2.0.14 and other versions.” Check yours with claude --version and run claude update if it is stale — an outdated binary can surface a model-doesn’t-exist error that looks like a GLM problem but isn’t. If the config still doesn’t take, Z.ai’s own FAQ escalation is: open a fresh terminal again; if that fails, delete ~/.claude/settings.json entirely and reconfigure — Claude Code regenerates the file automatically — and check your JSON for missing or extra commas. An independent walkthrough published June 30, 2026 on the AI Maker newsletter confirms the same sequence works end to end, two weeks after launch.

The setup is the easy half; the routing decision is where teams actually win or lose money. The honest frame, supported by the vendor’s own published numbers: GLM-5.2 lands near-frontier on many single-shot coding benchmarks at a fraction of the cost, and trails Opus 4.8 on sustained long-horizon agent work. The chart below is drawn from Z.ai’s release blog on Hugging Face — vendor-stated scores, cited as such.

The pattern in those five rows is not random — it is the trend that should shape your routing. On short-feedback-loop tasks, where the model proposes and a human or test suite verifies within minutes, open-weight models have effectively converged with frontier. Where the gap survives is compounding autonomy: every extra unsupervised step multiplies small quality differences, which is why the deficit grows from four points on Terminal-Bench to a factor of two on SWE-Marathon. The moat has moved from single answers to sustained judgment. For the deeper score-by-score analysis, see the full benchmark breakdown.

Community enthusiasm for the swap is real, and worth hearing in the original voice — with the caveat that it describes exactly the short-loop work the table routes to GLM:

Treat that as one practitioner’s read on short-loop tasks, not a blanket equivalence — the long-horizon rows above are where the characterization stops holding. If you want this thinking as a general discipline rather than a one-model decision, a broader model-routing framework covers scoring, fallbacks, and evals across providers — and our AI transformation engagements typically start by building exactly this routing layer against a client’s own repos.

Every comparison post quotes the Opus number, because it is the flattering one. But if your daily driver inside Claude Code is the Sonnet slot — and for most developers it is — the honest multiple is much smaller. GLM-5.2’s list pricing is $1.40 per million input tokens and $4.40 per million output tokens, with cached input at $0.26 (the pricing page currently marks cached-token storage as limited-time free — vendor-stated). Here is what that means per slot:

Two readings follow. First, if you mostly run the Sonnet slot, the savings today are modest — roughly 1.4× on input and 2.3× on output during Sonnet 5’s intro window — and a Coding Plan subscription may beat pay-as-you-go entirely for steady daily use. Second, the picture shifts on September 1, 2026, when Sonnet 5’s announced list rates step up to $3 / $15 and the same swap widens to roughly 2.1× and 3.4×. Prices on both sides of this table have already moved once this year; treat the multiples as a July 2026 snapshot and re-run the division before making a platform bet. If you’d rather buy GLM-5.2 tokens somewhere other than Z.ai first-party, our companion guide compares where to buy GLM-5.2 API access across hosts.

For sustained daily coding, most people will run this through a GLM Coding Plan rather than per-token billing. Three tiers, at list pricing:

The quota mechanics reward timing. Per the devpack overview, GLM-5.2 prompts draw down quota at a 3× multiplier during peak hours — defined as 14:00–18:00 China Standard Time (UTC+8), which is 06:00–10:00 UTC, early morning in Europe and overnight on the US East Coast — and 2× off-peak, with Z.ai advertising a temporary 1× off-peak promotion through the end of September 2026 (vendor-stated). Western working hours mostly fall off-peak, which quietly stretches every tier. Z.ai’s subscribe page also lists billing-cycle discounts — deeper for quarterly and yearly commitments, with a yearly Lite working out near $12.60 a month — and some Z.ai surfaces display already-discounted monthly prices, so treat $18 / $72 / $160 as the list anchor. For a full usage-pattern breakdown of whether the $18 plan is worth it for your usage pattern, we ran the numbers separately.

If the tier math fits your workload, you can subscribe through our GLM Coding Plan link — per Z.ai’s campaign rules, the discount applies to a new, never-paid account’s first subscription order only, doesn’t stack with other first-order campaigns, and requires completing payment within 72 hours of clicking through.

Referral link: we earn Z.ai platform credits if you subscribe, and new Z.ai accounts get 10% off their first subscription order.