把 Jev 当路由层,成本降到旗舰模型的 1/13
作者用 Jev 做前置路由判断任务是否需要大模型,60 道难题上准确率追平旗舰模型,成本约为其 1/13。
Jev is really f*ckin cheap it is not a chatbot. it does not write text. that is the whole new model. you give it a situation and a question, and it just picks an answer - yes/no, one option from a list, or a score - and tells you how confident it is. because it never writes anything back, you only pay for what goes in. ~$0.04 per million tokens in, output is free. a normal model bills you for every word it generates. here is the move from the article: stop making the expensive model answer everything. put Jev in front as a router. for each task Jev decides one thing - "does this actually need the big model?" easy ones go to the cheap model, only the hard ones get escalated to the flagship. on 60 hard math problems that combo scored the same as the $20 flagship - at about 1/13 of the price. the always-cheap model on its own lost 32 points. frontier accuracy, discount bill. that is it.