1) you can parallelize the request 2) since a structured output is still just text then every previous property of the structured output effects subsequent ones

Mostly, they are much faster and cheaper.