Being able to force JSON output for an LLM is huge for developers - 95% of my experiments with other LLMs for product ideas end when I can’t get the output in the format I need. GPT-3 and up were already the best at this - not having to even waste tokens on it is even better.