Hi guys! Has anyone managed to do real LLM routing to optimize costs on OpenClaw? I've given it instructions. I've added them to its soul.md file, created sub-agents that are supposed to use the right model depending on the complexity or type of task, but nothing works. It always uses the default (“primary”) model saved in its config (onclaw.json). Because there are quite a few videos out there, the guys claim to have reduced token costs by 80, 90, 97% in their OpenClaw, but I get the impression that it's all a myth. For now, the only option I have is to create a workspace for each agent with its own default model. This would apparently involve having to restart the gateway or session to switch agents, and they would each live in two different worlds (workspaces) that are completely disconnected with no possibility of communicating with each other. If anyone has a method that really works, I'm interested. For now, I'm 100% on Kimi K2.5. I don't give it ultra-complex tasks, so it does the job for not too much money.

sooo I set up OpenClaw this weekend and made the rookie mistake of running everything through Claude Opus. $75+ a day just chatting with the thing. Then I realized most of what I was asking it didn't actually need the smartest model. So I set up model routing instead: Gemini Flash for everyday stuff, Sonnet for writing, Opus only for complex builds. If you're using one model for everything, try splitting by task complexity. My daily cost dropped to about $5 overnight. Literally just tell OpenClaw to set this up for you.