Model Routing and Fusion Are Becoming Real Infrastructure Patterns
Two threads today cover model routing and fusion: Sakana's Fugu project and a separate mention of OpenRouter's fusion API. The core idea is that instead of picking one model, you route queries to whichever model handles them best, or fuse outputs from multiple models to get better answers. Commenters immediately draw comparisons to OpenRouter, which is already doing something similar.
The GLM 5.2 vs. Claude Opus thread is adjacent to this. People are not just asking 'which model is best' anymore; they're asking 'which model is best at what price for which type of query.' That's a routing question, not a ranking question. GLM-5.2 is less than a fifth the cost of Opus on output tokens, which makes it a serious option for high-volume tasks even if it loses on raw capability benchmarks.
The pattern: the AI infrastructure layer is maturing from 'pick a model' to 'build a routing layer that picks the right model per task.' This is a real engineering problem and a real product opportunity.
So what?
If you're spending more than a few hundred dollars a month on API costs, a routing layer that sends cheap queries to cheap models and hard queries to expensive ones could cut your bill by half or more. GLM-5.2's pricing versus Opus is a concrete example. This is also a competitive moat: builders who get good at model routing will outcompete those who just default to the most expensive model for everything.