The drift problem is real and I've found it gets 10x worse when you add AI/LLM endpoints to the mix — those change constantly as you iterate on prompts and response schemas. What's worked best for me: treat the OpenAPI spec as a build artifact that's auto-generated from code annotations. If the spec is handwritten separately from the code, it WILL drift. Tools that auto-sync are great, but the root fix is making documentation a byproduct of development, not a separate step. For teams building AI APIs specifically, I'd also recommend versioning your prompt templates alongside your API versions — that's a documentation gap most teams don't think about until it's too late.