Structured Outputs Make JSON Parsing Optional

My take: the "parse later" pattern is turning into legacy plumbing. The hard part is no longer cleaning up malformed JSON, it is deciding where to enforce structure and how much semantic risk to tolerate.

• –OpenAI and Gemini now support schema-based structured outputs, which cuts down retries, brittle regex fixes, and post-processing code.
• –Self-hosted stacks are catching up too, with constrained-decoding engines like XGrammar and llguidance making grammar-enforced generation practical.
• –Removing JSON repair does not remove validation; it shifts the work to schema design, versioning, refusals, and business-rule checks.
• –Too much structure too early can still hurt output quality, so many teams will want a hybrid flow: reason freely, then constrain the tool or API handoff.