Used to spend mental energy every single prompt deciding: should this go to Claude? Or is DeepSeek good enough? GPT-4o maybe?
After a month of doing this manually, I automated it. Now a routing layer classifies each task and picks the model for me.
What Auto-Selection Actually Looks Like
I send a task. The router looks at it and decides:
- File read? → DeepSeek (fast, cheap)
- Code refactor? → DeepSeek if simple, Sonnet if multi-file
- Code review? → GPT-4o (catches different things)
- Architecture decision? → Claude Sonnet (nothing else matches)
- Summarization? → Gemini Flash (fastest)
I never think about model selection anymore. I just work.
Why This Matters More Than Cost Savings
Yes, it saved me ~40% on my API bill. But the bigger win is cognitive load. Every time you decide which model to use, that is a micro-decision that drains focus. Over 50+ tasks per day, it adds up.
Now my workflow is: write prompt → get result. The model selection is invisible.
How I Set It Up
I use TeamoRouter for this. One API key, install in OpenClaw in 2 seconds. The teamo-balanced mode handles the auto-selection.
The routing is not perfect — maybe 5% of the time it picks a cheaper model when I would have preferred Sonnet. But I can always override with teamo-best for specific tasks.
The Unexpected Benefit
Rate limits basically disappeared. Since my requests spread across 4 providers instead of hammering one, no single provider sees enough traffic to throttle me.
We have a Discord where people share their routing configs. Curious how others handle model selection — manual or automated?
Top comments (0)