Have they ever said what the difference is between Sonnet and Opus? Are they trained differently? Different architectures? Is Sonnet a distillation? Is it just that Sonnet has less resources for inference?

None of the other labs are doing this kind of long lived two model series.