How I connected OpenClaw to a Chinese frontier model and cut agent costs by 90%

I've been running OpenClaw for a few months. Love the agent workflow. Not so much the monthly bill.

Last week I set up a tiny routing layer that lets OpenClaw fall back to DeepSeek V4 for certain task types. The config change was 4 lines.

Here's what happened:

• Code review tasks: ~$0.15 → ~$0.002 each
• Batch processing: costs dropped from $20/day to ~$2/day
• The agent didn't notice the difference — same structured output, same format

The setup is straightforward. You need any OpenAI-compatible endpoint that supports the models you want. Drop a provider block into your OpenClaw config, swap the model name, restart.

Not claiming this replaces Claude for complex reasoning. But for high-volume structured work, having a cheap fallback available through the same interface changes how aggressively you can let the agent run autonomously.

Happy to answer questions about the routing setup.