This post matches my experience. You get what you pay for and the performance is so uneven that you end up having to tightly pin providers. Those pinned providers then have very little skin in the game. I agree with the unstated conclusion here - I don’t think OpenRouter is fit for production workloads today. It’s great for scrappy/quick/cheap validation though as you’re dialling in on taste and want to quickly try many different open models (Kimi, GLM, DS etc.). Know your tools.