# Use BigModel.io with Claude Code

> Point Claude Code at the BigModel.io Anthropic-compatible API with ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN, pick GLM, Kimi, DeepSeek or MiniMax models, and fix common errors.

Web version: https://bigmodel.io/docs/claude-code · Updated 2026-10-18

Claude Code talks the Anthropic Messages protocol. BigModel.io serves it at `https://bigmodel.io/v1/messages`, so you configure the gateway root and Claude Code adds `/v1/messages` itself. For OpenAI-protocol code, the base URL is `https://bigmodel.io/v1` instead.

| Setting | Value |
|---|---|
| `ANTHROPIC_BASE_URL` | `https://bigmodel.io` (no `/v1`) |
| `ANTHROPIC_AUTH_TOKEN` | your BigModel.io API key (`sk-...`), sent as `Authorization: Bearer` |
| `ANTHROPIC_MODEL` | a model ID from the [model catalog](https://bigmodel.io/models) |
| `ANTHROPIC_DEFAULT_HAIKU_MODEL` | a cheaper model for background tasks |

> Anthropic states that Claude Code does not officially support non-Claude models through a gateway. GLM, Kimi, DeepSeek and MiniMax models work through BigModel.io, but Anthropic will not help with problems specific to them, and quality on long agentic tasks varies by model.

## Before you start

1. Install Claude Code and confirm `claude --version` works.
2. Create a key in [API Keys](https://bigmodel.io/keys). Its group decides which models and endpoints you can call.
3. Check that the key's group serves the Messages endpoint. GLM and DeepSeek groups that are "Chat Completions only" return 404 on `/v1/messages`; for those keys, use an OpenAI-protocol tool such as [OpenCode](https://bigmodel.io/docs/opencode) or [Cline](https://bigmodel.io/docs/cline).

This request should return a short reply. Replace `glm-5.3` with the model you plan to use:

```bash
curl https://bigmodel.io/v1/messages \
  -H "Authorization: Bearer $BIGMODEL_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model":"glm-5.3","max_tokens":64,"messages":[{"role":"user","content":"ping"}]}'
```

## Configure with environment variables

For a single terminal session:

```bash
export ANTHROPIC_BASE_URL="https://bigmodel.io"
export ANTHROPIC_AUTH_TOKEN="${BIGMODEL_API_KEY:?Set BIGMODEL_API_KEY first}"
export ANTHROPIC_MODEL="glm-5.3"
# Claude Code also runs small background tasks; point them at a model you can call.
export ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-5.3-flash"

claude
```

Use `ANTHROPIC_AUTH_TOKEN`, not `ANTHROPIC_API_KEY`. `ANTHROPIC_API_KEY` is sent as `x-api-key` and needs a one-time approval in interactive mode; if both are set, Claude Code reports an auth conflict. Run `unset ANTHROPIC_API_KEY`.

## Configure in settings.json

To keep the setting across sessions, add an `env` block to `~/.claude/settings.json` (all projects) or `.claude/settings.local.json` (one project, not committed):

```json
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://bigmodel.io",
    "ANTHROPIC_AUTH_TOKEN": "sk-your-bigmodel-key",
    "ANTHROPIC_MODEL": "glm-5.3",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.3",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.3",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.3-flash"
  }
}
```

Never put the key in a committed `.claude/settings.json`. Values in a settings file win over a shell `export`. Claude Code's `opus`, `sonnet` and `haiku` aliases in `/model` map to the three `ANTHROPIC_DEFAULT_*_MODEL` values, and the haiku slot is also used for background tasks, so set all three to models your key can call. `ANTHROPIC_SMALL_FAST_MODEL` is deprecated.

### VS Code extension

The Claude Code extension for VS Code does not read the variables you `export` in a terminal. Either:

- put the `env` block above in `~/.claude/settings.json`, which the extension and the CLI share (recommended), or
- add the variables to the extension's `claudeCode.environmentVariables` setting in VS Code.

Also turn on **Disable Login Prompt** (`claudeCode.disableLoginPrompt`) so the extension does not ask for a claude.ai sign-in.

Claude Desktop ignores `ANTHROPIC_BASE_URL`, so it cannot be pointed at BigModel.io this way.

## Choosing a model

Model IDs are lowercase and keep their dots (`deepseek-v4-pro`, not `DeepSeek-V4-Pro`). Copy the ID from the model page; availability depends on your key's group.

| Family | Example `ANTHROPIC_MODEL` | Notes |
|---|---|---|
| GLM (Zhipu) | `glm-5.3` | `glm-5.3-flash` suits the haiku slot |
| Kimi (Moonshot) | `kimi-k3` | |
| DeepSeek | `deepseek-v4-pro` | `deepseek-v4-flash` is a cheaper fast option |
| MiniMax | `minimax-m3` | |

Switch inside a session with `/model glm-5.3`, or start with `claude --model kimi-k3`. Try a small task before a long agentic run.

### Context window for non-Claude models

Claude Code does not know these model IDs. It shows a warning that the model "isn't recognized" and assumes a 200k-token context. Set `CLAUDE_CODE_MAX_CONTEXT_TOKENS` to the model's real context window (from its model page) so automatic compaction starts at the right time:

```bash
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=128000   # replace with the model's context window
```

Do not enable `CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY`: it keeps only model IDs that contain `claude` or `anthropic`, so BigModel.io models will not appear. Use `ANTHROPIC_MODEL` and the alias variables instead.

### Other useful variables

| Variable | When to set it |
|---|---|
| `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1` | You get 400 "Extra inputs are not permitted" or a rejected `anthropic-beta` header |
| `API_TIMEOUT_MS` | Long requests time out; the default is `600000` (10 minutes) |
| `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1` | Optional; turns off non-essential traffic to Anthropic |

## Verify the setup

Run `/status` inside Claude Code. It should show `https://bigmodel.io` as the Anthropic base URL and an auth token line from your environment or settings. Then send one prompt and confirm the request appears in [Usage](https://bigmodel.io/usage). A reply alone does not prove the right base URL was used.

## Troubleshooting

**403 `permission_error`: `This group does not allow /v1/messages dispatch`.** The key's group doesn't offer the Anthropic Messages endpoint. Use a key from a group that supports it, or connect your tool through the OpenAI-compatible endpoint instead ([OpenCode](https://bigmodel.io/docs/opencode), [Cline](https://bigmodel.io/docs/cline)). [Contact support](mailto:support@bigmodel.io) if you're unsure which group to use.

**404 `not_found_error`: `... platform only supports /v1/chat/completions; /v1/messages is not available`.** The key belongs to a Chat Completions-only group (GLM or DeepSeek platform). Use a key from a group that serves Messages, or use an OpenAI-protocol tool.

**404 on every request.** Check `ANTHROPIC_BASE_URL`. With `https://bigmodel.io/v1` the client calls `/v1/v1/messages`. Remove the `/v1`.

**401 `INVALID_API_KEY` or `API_KEY_REQUIRED`.** The token is missing or mistyped, or Claude Code is still using a claude.ai login. Check `/status`, make sure `ANTHROPIC_AUTH_TOKEN` is set where `claude` starts, and remove any `ANTHROPIC_API_KEY`.

**503 `api_error`: `No available accounts: no available accounts supporting model: ...`.** The model ID is misspelled, or your key's group does not serve it. List what your key can call with `curl https://bigmodel.io/v1/models -H "Authorization: Bearer $BIGMODEL_API_KEY"`.

**400 "Extra inputs are not permitted".** Claude Code sent a beta field or `anthropic-beta` header that the model path rejects. Set `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1` and restart.

**Warning "isn't recognized · context assumed to be 200k tokens".** Expected for non-Claude models. Set `CLAUDE_CODE_MAX_CONTEXT_TOKENS` as shown above.

**Main replies work but titles, summaries or subagents fail.** Background tasks use the haiku slot. Set `ANTHROPIC_DEFAULT_HAIKU_MODEL` to a model your key can call.

**403 `INSUFFICIENT_BALANCE` or 429 `API_KEY_QUOTA_EXHAUSTED`.** Top up in [Billing](https://bigmodel.io/purchase) or raise the key's quota in [API Keys](https://bigmodel.io/keys). A 429 `rate_limit_exceeded` with a message like "api key 5小时限额已用完" means a key spend window is used up; it resets when the window ends.

**409 `malformed_session_state`.** The conversation history has thinking blocks or tool IDs the model cannot accept, often after switching models mid-session. Start a new session with `/clear`.

**Settings seem ignored.** A settings file beats a shell `export`, and a project `.claude/settings.json` or `.claude/settings.local.json` can override your user file. Run `env | grep ANTHROPIC`, check the project's settings files, and restart Claude Code after changes.

## Related

- [Configure Base URL in Codex, Claude Code, Cursor and Cline](https://bigmodel.io/blog/configure-base-url-codex-claude-code-cursor) — the same settings side by side.
- [CC Switch](https://bigmodel.io/docs/cc-switch) to import a key into Claude Code with one click.
- [Codex CLI](https://bigmodel.io/docs/codex-cli) and [OpenCode](https://bigmodel.io/docs/opencode) for OpenAI-protocol terminal agents.
- [Errors](https://bigmodel.io/docs/errors) for every error code.
- [API quickstart](https://bigmodel.io/docs/quickstart) for keys, billing and first requests.