That's... not always a given for SOTA sized models. When the ROI on more training stops, it is nice to have alternatives, whether that is RL-tuned reasoning models or alternative architectures that improve specific areas of weakness.