“In the long term, all large language models will be reasoning models, and efficiency gains will make them practical for all use cases.”