Hacker News Reader

Is there an architectural reason Sonnet beats 4o? Or is it simply a matter of training corpus?

We don't know what exact architecture and optimizations both of those models use, it's all proprietary nowadays :)