aw models
Open chat ↗

Claude Code
on Raw Models

Claude Code sends every request to one API base URL, so pointing it at Raw Models replaces its models entirely: the native Claude models are not reachable while this is set, and each tier maps to a Raw Models model of your choice. Reverting restores them.

One command

Maps every tier to a single model (default glm-5.3-raw; override with RAWMODELS_MODEL). Create a key under settings → API keys.

curl -fsSL https://www.rawmodels.ai/install | RAWMODELS_API_KEY=skrm-... sh -s -- claude

Restart claude, then run any model by slug, for example:

claude --model glm-5.3-raw -p "do your job"

Restart claude afterwards. Undo any time: curl -fsSL https://www.rawmodels.ai/install | sh -s -- uninstall. On macOS the installer reads the key from the terminal; the security command in examples is macOS-only — on Linux just paste the key when prompted or pass RAWMODELS_API_KEY=.

Per-tier mapping (manual)

To give each tier a different model, add this to the env block of ~/.claude/settings.json (merge into an existing env if present), then restart Claude Code.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://www.rawmodels.ai",
    "ANTHROPIC_AUTH_TOKEN": "skrm-...",

    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.3-raw",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.3-raw",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.3-raw",
    "ANTHROPIC_MODEL": "glm-5.3-raw",
    "ANTHROPIC_SMALL_FAST_MODEL": "glm-5.3-raw",
    "CLAUDE_CODE_SUBAGENT_MODEL": "glm-5.3-raw",

    "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "64000",
    "API_TIMEOUT_MS": "900000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}
ANTHROPIC_DEFAULT_OPUS_MODEL

What the opus / opusplan tier resolves to.

ANTHROPIC_DEFAULT_SONNET_MODEL

The sonnet tier — the default working model.

ANTHROPIC_DEFAULT_HAIKU_MODEL

The haiku / small-fast tier, used for background work.

ANTHROPIC_MODEL

The model a fresh session starts on.

CLAUDE_CODE_SUBAGENT_MODEL

The model subagents run on.

CLAUDE_CODE_MAX_OUTPUT_TOKENS

Output budget per turn. Keep within the selected model’s limits.

Available models

Use a tool-capable model your account can access. Early access is required:

glm-5.3-raw

GLM-5.3 Raw — tool calling; early access required.

Uncensored models are coming soon. Join the early access waitlist. See pricing on the reference.

Revert

Remove the env block (or the keys above) from ~/.claude/settings.json and restart claude, or run the installer's uninstall. Your native Claude models come back.

← API reference