Unlock Faster, Cheaper Claude Code with Domestic LLMs: 3 Practical Solutions

Solution 1: DeepSeek (Native Anthropic‑compatible API)

DeepSeek offers a direct Anthropic‑compatible endpoint, allowing Claude Code to connect without a proxy.

Step 1 – Obtain API key

Register at https://platform.deepseek.com, create an API key, and select the "Claude Code" group.

Step 2 – Set environment variables (Linux/macOS)

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=YOUR_DEEPSEEK_API_KEY
export ANTHROPIC_MODEL=deepseek-reasoner
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

Windows PowerShell:

$env:ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
$env:ANTHROPIC_AUTH_TOKEN="YOUR_DEEPSEEK_API_KEY"
$env:ANTHROPIC_MODEL="deepseek-reasoner"
$env:CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

Step 3 – Verify

cd your-project
claude

The connection log should display deepseek-reasoner or deepseek-chat as the active model.

Key characteristics

Very low cost (near‑free for daily coding).

Strong reasoning capabilities, especially with deepseek-reasoner.

Low latency (< 1 s) due to domestic direct connection.

Good Chinese language support.

Solution 2: Alibaba Cloud Bailei + Tongyi Qianwen (Enterprise‑grade)

Bailei provides a commercial Coding Plan optimized for Claude Code and supports dedicated coding models.

Configuration (edit ~/.claude/settings.json)

{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "YOUR_BAILEI_API_KEY",
    "ANTHROPIC_BASE_URL": "https://coding.dashscope.aliyuncs.com/apps/anthropic",
    "ANTHROPIC_MODEL": "qwen3-coder-plus",
    "ANTHROPIC_SMALL_FAST_MODEL": "qwen-turbo"
  }
}

Model mapping recommendations

Daily coding – qwen3-coder-plus (comparable to Claude 3.5 Sonnet).

Fast completion – qwen-turbo (comparable to Claude 3 Haiku).

Complex architecture – qwen-max (comparable to Claude 3 Opus).

Measured performance (representative figures)

Code‑generation accuracy: DeepSeek 85 % vs Claude Sonnet 4.5 90 %.

Response latency: DeepSeek ≈ 400 ms vs Claude ≈ 2 s (requires VPN).

Cost: ¥0.6 per M tokens (DeepSeek) vs $15 per M tokens + VPN overhead (Claude).

Chinese variable‑name understanding: DeepSeek 95 % vs Claude 80 %.

Solution 3: Third‑party Relay Services (Flexible)

Relay services allow use of the native Claude model while bypassing network restrictions.

Typical providers

xingjiabiapi.org – lower price, latency 0.5‑1 s, success rate > 98 %.

claude-code.com.cn – supports latest Claude Opus 4.5/4.6, domestic nodes.

api.yixia.ai – simple one‑click setup.

Configuration example (xingjiabiapi.org)

export ANTHROPIC_BASE_URL=https://xingjiabiapi.org
export ANTHROPIC_API_KEY=sk-YOUR_KEY

Or edit ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_API_KEY": "sk-YOUR_KEY",
    "ANTHROPIC_BASE_URL": "https://xingjiabiapi.org",
    "ANTHROPIC_MODEL": "claude-opus-4-5-20251101"
  }
}

Advanced Tip: Model Auto‑Switching

Claude Code can automatically select a model based on task complexity; manual configuration provides finer control.

Recommended environment configuration

{
  "env": {
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "qwen-max",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "qwen3-coder-plus",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "qwen-turbo"
  }
}

Scenario mapping

File reading / simple syntax checks → qwen-turbo (fast, cheap).

Code generation / refactoring → qwen3-coder-plus (balanced).

Complex architecture design / deep reasoning → qwen-max (strong reasoning).

Quick switch command (enter in Claude Code)

/model qwen-max

Common Troubleshooting

Invalid API Key – Ensure the key starts with sk- and has the required permissions.

Failed to connect to api.anthropic.com – Add CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 and verify hasCompletedOnboarding: true in the config file.

Poor code‑generation quality – Adjust the model temperature to a range of 0.2‑0.5; lower temperatures improve stability for domestic models.

Multi‑model orchestration – Use LiteLLM to build a local gateway that aggregates multiple model APIs (requires Python).