Claude Code

Use ModelSite chat models in Claude Code: the three settings, the one thing to change for GLM models, and a one-line check

Status: run against ModelSite with Claude Code 2.1.309 on 2026-10-24 (claude -p, one request, answered correctly).

Settings

export ANTHROPIC_BASE_URL="https://api.modelsite.ai"
export ANTHROPIC_AUTH_TOKEN="$MODELSITE_API_KEY"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-5.3-flash"
export ANTHROPIC_DEFAULT_SONNET_MODEL="glm-5.3-flash"
export ANTHROPIC_DEFAULT_OPUS_MODEL="glm-5.3-flash"
claude --model glm-5.3-flash

Use ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY. The platform reads the key from Authorization: Bearer only; ANTHROPIC_API_KEY is sent as x-api-key and returns 401.
Claude Code uses its small haiku alias for background work. Point the three aliases at ModelSite model IDs as above, or those requests ask for models that do not exist here.
Claude Code prints a notice for a model ID it does not recognize and assumes a 200k context window. This is harmless.

GLM models and the effort level

The GLM models (glm-5.3, glm-5.3-flash) support the effort levels low, high and max. When the effort level is set to medium (for example with CLAUDE_CODE_EFFORT_LEVEL=medium or --effort medium), ModelSite runs the request at high. Other chat models accept every level as sent. You do not need to change any setting.

To pick a level explicitly:

export CLAUDE_CODE_EFFORT_LEVEL=high   # every session
claude --model glm-5.3-flash --effort low   # one session

Check it works

claude -p "Reply with exactly one word: pong" --model glm-5.3-flash --max-turns 1

You should see pong. A Claude Code request carries a long system prompt (about 9,500 input tokens in our run), so even this one-word check bills a small, non-zero amount.

Limits

Tool calls and long multi-turn sessions were not part of this check.
For other coding assistants (Codex CLI, OpenCode, Aider, Cline, Continue), see SDKs, Frameworks and Clients.

Settings

GLM models and the effort level

Check it works

Limits

On this page