Token efficiency with structured output from language models