What are other OpenClaw users seeing? Anyone found effective ways to cut costs (model tiering, caching, etc.)?