Hanselminutes with Scott Hanselman · Hands-On Engineering Podcasts · November 2026
Scott Hanselman asked whether OpenRouter is about finding the smartest model or getting the task done. Emma Marie Cryer warns that a cheap model that can't finish the job can cost more than a frontier model that can.
Like
…smartest thing possible. I feel like that's the wrong question. Like, I want to get the thing done. I want to be responsible in my use of compute, responsible in my use of inference. Do you think that open router is about finding the smartest model, or is it more about achieving the task at hand?
…completely transparent. I think it is more important, though, to think, you know, when you are completing a task, are you optimizing for intelligence? Are you optimizing for latency? Are you optimizing for speed? And if someone isn't thinking about that, that's not something that we can necessarily make them think about. But I do think it is important to think about in terms of token spend is something that I see people making the mistake of all the time is thinking about cost per token instead of cost per task. And that can burn a lot of tokens unnecessarily. If you have something you're working on that's a little more complex and you're trying to do it with open source, it might actually be a waste of resources because it could end up, if it's not a model with the level of capability you need, able to finish the task as quickly and efficiently as if you had just used a frontier model in the first place. So, you know, what model is the smartest?…
Yeah, you know, that's actually, I like that you kind of scratched a neuron in my brain there. I'm, this is random, but yesterday I borrowed a friend's car. It's a stick shift car, and I'm teaching my wife how to drive stick. Now she's driven for 30 years, so she knows how to drive. She doesn't…
Yeah, exactly. But that's not to say that open weight models are not really something that we should be leaning towards too. We released something called Fusion a while ago. And it was basically a set of benchmarks we ran where we had a bunch of open models together and then a bunch of frontier…
Yeah, that's a really good point. And again, this podcast is my side hustle, but in my day job, I do work at GitHub and they have a thing called HydroFusion. I think this concept of fusion, and you can See this at openrouter.ai/slash fusion. You can try it yourself. It's becoming really…