TBH, my biggest gripe with Google is that they seem to support a slightly different JSON schema format for structured outputs than everybody else. Where Open AI encourages (or even forces) you to use refs for embedding one object in another, Google wants you to embed directly, which is not only wasteful but incompatible with how libraries that abstract over model providers do it.

My structured output code (which uses litellm under the hood, which converts from Pydantic models to JSON schemas), does not work with Google's models for that reason.