How to Use GLM-5.2 in Claude Code with 1M Context

Independent research — not an official Z.ai publication.Identity and provider disclosure

AI Infrastructure Editor

Editorial responsibility: GLM52.ai Editorial TeamPublished July 21, 2026 · Editorially reviewed July 26, 2026

How we research and review Report a correction

Disclosed editorial pen name · AI-generated portrait

Sanitized Claude Code 2.1.220 test summary showing a successful GLM-5.2 Coding Plan request, Read tool, and incompatible metered route

To use GLM-5.2 directly in Claude Code, connect an active GLM Coding Plan to Z.ai’s Anthropic-compatible endpoint, map Sonnet and Opus to glm-5.2[1m], then verify a request and a harmless Read call. The pinned response returned CLAUDE_PLAN_OK in 2.648 seconds; the tool probe returned the fresh file sentinel in 3.613 seconds.

The required pay-as-you-go scenario is also tested and documented here, but its result is negative: the global Z.ai metered endpoint is not a direct Claude Code gateway. Use a client that speaks OpenAI-compatible Chat Completions or a deliberately operated translation gateway; do not relabel a 404 as an authentication problem.

In this guide

Confirm the access route
Add the Coding Plan settings
Test the pay-as-you-go boundary
Keep the Z.ai key outside the project
Verify the model, response, and Read tool
Budget the 1M context window
Choose high or max effort
Fix common failures
Know what this setup proves
Sources and method

Confirm the access route

Z.ai’s model-switching guide defines this setup for existing GLM Coding Plan users. The plan is designed for supported interactive coding agents. It is not ordinary pay-as-you-go API credit, and its endpoint is not the OpenAI-compatible URL used by OpenCode, Cline, or an application backend.

Client or job	Protocol	Documented endpoint or model ID	Use this page?
Claude Code with Coding Plan	Anthropic-compatible	`https://api.z.ai/api/anthropic`	Yes
OpenCode with Coding Plan	OpenAI-compatible	`https://api.z.ai/api/coding/paas/v4`	Use the OpenCode guide
Claude Code with pay-as-you-go API	Anthropic client against OpenAI Chat route	`https://api.z.ai/api/paas/v4`, model `glm-5.2`	Direct test failed with HTTP 404
General pay-as-you-go application	OpenAI-compatible	`https://api.z.ai/api/paas/v4`, model `glm-5.2`	Use the API provider guide

This distinction prevents a frustrating error class: a valid key sent to the wrong product or protocol can still fail. If you have not chosen an access route, compare Coding Plan, API, and self-hosting before editing Claude Code.

Add the Coding Plan settings

Claude Code documents user settings at ~/.claude/settings.json on macOS, Linux, and WSL, or %USERPROFILE%\.claude\settings.json on native Windows. Preserve the rest of the file. Add or update only the keys inside env:

{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-4.5-air",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.2[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.2[1m]",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "API_TIMEOUT_MS": "3000000"
  }
}

The model aliases matter because Claude Code can assign different work to its Haiku, Sonnet, and Opus slots. This map keeps the lightweight slot on glm-4.5-air and routes the two main slots to the long-context GLM-5.2 selector. Z.ai’s Claude Code page publishes the same mapping.

Do not replace a complete settings file with this example. A real file may also contain permission rules, hooks, plugins, or an organization policy. JSON objects can be merged; duplicate top-level keys, missing commas, and a second env object cannot.

You can syntax-check a copy without exposing a credential:

node -e 'JSON.parse(require("node:fs").readFileSync(process.argv[1], "utf8")); console.log("valid JSON")' \
  "$HOME/.claude/settings.json"

Syntax is only the first gate. Claude Code can load settings from managed, command-line, local, project, and user scopes. User settings have the lowest precedence. A project or local model override can therefore win even when this JSON is valid.

Test the pay-as-you-go boundary

The global metered route cannot be made compatible merely by replacing the URL and removing [1m]. We preserved and executed this negative probe:

{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/paas/v4",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.2",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.2",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.2",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "API_TIMEOUT_MS": "300000"
  }
}

Claude Code 2.1.220 sent an Anthropic Messages-shaped request and received HTTP 404 after 1.589 seconds. Z.ai’s general API documentation exposes Chat Completions at this route. That is a protocol mismatch, not evidence that the model ID, key, or account is invalid.

For metered GLM-5.2 use, the simpler answer is OpenCode, Cline, Kilo, Crush, goose, Hermes, OpenClaw, Pi, or a direct application client. A translation gateway can theoretically convert protocols, but it adds data handling, feature-loss, logging, and operational risks; this article did not verify such a Claude Code gateway and does not label one as working.

Keep the Z.ai key outside the project

Z.ai’s example accepts ANTHROPIC_AUTH_TOKEN. Set it through your shell’s protected secret loader or credential manager. For a temporary current-shell test:

export ANTHROPIC_AUTH_TOKEN="your-zai-key"
test -n "$ANTHROPIC_AUTH_TOKEN" && echo "Z.ai token is set"

$env:ANTHROPIC_AUTH_TOKEN = "your-zai-key"
if ($env:ANTHROPIC_AUTH_TOKEN) { "Z.ai token is set" }

The presence checks do not print the secret. Do not use echo $ANTHROPIC_AUTH_TOKEN, put the key in .claude/settings.json inside a repository, paste it into an issue, or include it in a screenshot. Anthropic’s gateway documentation says this token is sent as authorization to the configured base URL, so treat it as a credential, not a model preference.

If a key has entered shell history, source control, a screenshot, or support text, revoke it in the provider console. Deleting the visible copy does not make an exposed credential safe again.

Verify the model, response, and Read tool

Start a fresh terminal so it inherits the intended environment, then identify the installed client:

claude --version
claude

Inside Claude Code, run:

/status

Require two visible facts:

Setting sources includes the user file you edited, normally ~/.claude/settings.json.
Model shows glm-5.2 or glm-5.2[1m] for the active main slot.

Both Z.ai and Claude Code recommend /status for this check. Asking the model “what model are you?” is weaker evidence: a generated answer can repeat a system label or guess. The rendered status is client state.

After the status check, use a bounded, read-only task in a disposable directory before granting write or shell permissions:

printf 'CLAUDE_GLM52_%s\n' "$(date +%s)" > claude-glm52-check.txt

Ask Claude Code to use Read on that file and return only its contents. A valid result contains a visible Read event and an exact match to the freshly generated token.

The pinned Docker evidence recorded:

Claude Code 2.1.220 check	Result
Coding Plan response	`CLAUDE_PLAN_OK`, 2,648 ms wall / 2,270 ms API
Response tokens	124 input, 25 output; model `glm-5.2[1m]`
Coding Plan `Read`	`GLM52_TOOL_PROBE_20260726` returned, 3,613 ms wall / 3,596 ms API
Tool-call tokens	230 input, 59 output, 1,856 cache-read
Direct metered route	HTTP 404, protocol incompatible

A successful read proves one request and one built-in tool round trip; it does not prove long-session reliability, million-token recall, writes, shell execution, or every MCP server.

Budget the 1M context window

Z.ai documents two required settings for the long-context Claude Code path: the glm-5.2[1m] selector and CLAUDE_CODE_AUTO_COMPACT_WINDOW set to 1000000. The GLM-5.2 model guide separately publishes a 1M context and 128K maximum output.

That capacity is not a promise that Claude Code can paste one million project tokens into a clean prompt. System instructions, tool definitions, existing messages, tool results, selected files, and output headroom share the window.

Use this worksheet:

project material
= 1,000,000 auto-compact window
− instructions and tool definitions
− existing conversation and tool results
− planned output allowance
− safety margin

An illustrative allocation is:

1,000,000 − 70,000 − 120,000 − 64,000 − 50,000 = 696,000 tokens

The 70K, 120K, 64K, and 50K deductions are planning inputs, not measured Claude Code constants. Replace them with your own session counts. The arithmetic adds independent value because it exposes the real decision: every extra file competes with history, tools, output, and margin.

A five-percent margin in the example absorbs estimation error and late tool results. It does not guarantee acceptance. A plan tier, intermediary, client version, or account policy may expose a lower practical limit. A large accepted prompt also says nothing about recall quality; verify required facts and citations at the beginning, middle, and end of representative repositories.

Choose high or max effort

Z.ai documents a many-to-two mapping for Claude Code’s /effort command:

Claude Code selection	GLM-5.2 effort used by Z.ai
`low`, `medium`, `high`	`high`
`xhigh`, `max`, `ultracode`	`max`

This means moving from low to high does not select three distinct GLM levels on this route. Use high as the clear default label. Use max when a difficult refactor, multi-step diagnosis, or expensive failure justifies more reasoning and quota use.

Do not turn the mapping into a quality guarantee. It is a publisher-documented routing rule, not an independent benchmark. For the direct API’s separate reasoning_effort values and cost tradeoff, use the GLM-5.2 reasoning-effort guide.

Fix common failures

`glm-5.2[1m]` does not exist

Run claude update, open a new terminal, and check /status again. Z.ai explicitly lists an outdated Claude Code version as the first fix. Do not silently remove [1m] and continue to claim a one-million-token configuration.

`/status` shows the wrong settings source

Confirm that you edited the file used by the Claude Code installation you launched. Native Windows, WSL, and Git Bash can resolve different home directories. Then inspect project and local settings: Claude Code gives them higher precedence than user settings.

The request returns 401 or 403

Confirm token presence without printing it, confirm the account has an active Coding Plan, and confirm ANTHROPIC_BASE_URL is exactly https://api.z.ai/api/anthropic. A general API key or OpenAI-compatible endpoint is not interchangeable with this plan route. Repeated retries do not fix a product or credential mismatch.

The metered `/paas/v4` route returns 404

Do not rotate the key first. This was the expected result in our direct pay-as-you-go probe: Claude Code’s Anthropic Messages request does not match the global route’s OpenAI Chat Completions protocol. Choose a compatible client or an independently audited gateway.

Context compacts early or the server rejects the request

Check all three boundaries: the [1m] model selector, the 1000000 auto-compact setting, and the provider/account limit. Then reduce selected project material and reserve explicit output headroom. The prompt-caching guide explains why caching can lower repeat-input cost without enlarging context.

Tools fail although plain text works

Separate transport from agent behavior. First require one short response. Then require one harmless read. Only then test edits, shell commands, MCP, subagents, and long sessions. A text response proves authentication and response parsing once; it does not prove every Claude Code tool schema or permission path.

Images appear to work

GLM-5.2 itself is text input/text output in the cited model guide. A Claude Code image workflow can involve a separate vision or MCP component. The GLM-5.2 image-support guide shows how to trace which component handled the pixels.

Know what this setup proves

The documented settings and live run prove one direct Coding Plan route. A valid /status proves that Claude Code loaded a settings source and selected a model name. The exact response plus sentinel read prove one request and one tool exchange at one time. The 404 separately proves that the tested direct pay-as-you-go URL was not compatible with Claude Code’s protocol.

None of those prove perfect one-million-token recall, a fixed plan quota, stable latency, model identity at every intermediary, or production reliability. Record the Claude Code version, endpoint class, model label, effort, task, tool permissions, context estimate, timestamp, and visible result when a decision depends on the test.

Sources and method

Z.ai Claude Code guide — endpoint, environment variables, model mapping, and /status flow.
Z.ai model-switching guide — [1m], auto-compaction, effort mapping, and missing-model fix.
Z.ai GLM-5.2 guide — input/output modality, context, maximum output, and general API model ID.
Z.ai Coding Plan overview — subscription and supported-tool boundary.
Claude Code settings — settings scopes, precedence, validation, and status inspection.
Claude Code LLM gateway guide — custom base URL, authorization token, and provider model routing.

The article was rechecked on July 26, 2026. The live evidence used Claude Code 2.1.220 in a disposable non-root Docker container with a read-only root filesystem and no writable host mount. It called both product URLs with their intended credential context, preserved only sanitized measurements, and removed the task environment afterward. The results remain bounded to one account, one version, and one point in time.