Qwen, DeepSeek, GLM & Kimi Users - Real-Work Task Outcomes

27 auto-recharge receipts from Anthropic hit my inbox last weekend, all from one project. I rebuilt my portfolio site and handed a frontier model every single task, from the big architecture decisions down to formatting fixes. But, I knew better. Smart routing was half the talk at the AI Engineer World's Fair a couple of weeks ago. Speakers advised on using cheap models for the menial work, frontier models for the judgment calls. I nodded along, then kept giving everything to the Fable and Opus anyway. This weekend's receipts finally pushed me over the edge. I want to be clear: this isn't an Anthropic complaint. I love the products. I created this routing problem, and this week I'm fixing it in public. My experiment: I'm re-running some of my toil tasks on Chinese open-weight models, Qwen through a cheap API and then locally on my MacBook Air. The frontier model keeps what it's genuinely best at: architecture calls, tradeoffs, final review. Everything repetitive, the edits and formatting and scaffolding, goes to the cheaper tier. I am hoping for actual savings, but I'm also really interested in where the cheap models fail. I want to map what breaks when you go thsi route, and I plan on publishing my findings early next week. Here's where I need your comments: If you've run real work on Qwen, DeepSeek, GLM, or Kimi, tell me the one task you handed off and whether it survived. What broke, what held up, what it saved you. #AIEngineering #LLMOps #OpenWeightModels