Thanks for sharing your experience. We also get pretty reliable JSON output (on a smaller scale though) even without JSON mode. We usually don't use JSON mode

Thanks for sharing your experience. We also get pretty reliable JSON output (on a smaller scale though) even without JSON mode. We usually don't use JSON mode because we often include a chain of thought part in <brainstorming> and then ask for JSON in <json> tags. With some prompt engineering, we get over 98% valid JSON in complex prompts (with long context and modestly complex JSON format). We catch the rest with json5.loads, which is only used as a fallback if json.loads fails. 4o-mini has been less reliable for us particularly with large context. The new structured output might make it possible to use mini in more situations.