Has anyone published accuracy and speed benchmarks for BAML vs structured outputs from current models like gpt-5 and gemini-3? I came across boundary's original blog post from a year ago but it doesn't seem to have been updated with comparisons with newer models' native structured output capabilities. I would also be very curious to see a comparison when the output schema is recursive (e.g. context-free grammar) rather than simpler fixed-depth schemas.

We actually redid our benchmarking a few months ago during a hackathon, but didn't quite get around to writing it up and publishing it. Thanks for the push there. tl;dr we were seenig similar numbers to what we published a year ago. The hackathon work included a shim app that lets us run any query against both BAML and OpenAI tool calling, so we would be in a position to run your specific recursive structured output benchmark with that. Great point!

@greg (BAML) is the hackathon app on github? I'd love to run some comparison tests myself

It is! https://github.com/BoundaryML/openai-baml-adapter But we haven't polished it up enough to announce it - it may be quite buggy!