I've been on a client consulting project since mid-March that uses LLMs to generate SQL. Over the last 6 months, the accuracy of LLM-generated advanced SQL (think 1000–1500 lines in one massive query, with multiple joins, window functions, aggregations, and recursive CTEs) has improved a lot in our internal benchmark. The spider2-sql.github.io leaderboard shows the same trend. Google hasn't given us access to Gemini-SQL2 yet, so I can't speak to that one. But overall Gemini 3.1 is simply the best, no matter the database and claude just sucks - at least from what we see. I'm from the pre-ORM generation. We wrote SQL by hand, and DBMS with a lab was a college subject, so almost everyone learned it. Still, writing highly optimised queries was always an acquired skill. Every org had a few senior devs who could write or debug those big complex queries, and it was a no-go zone for everyone else. You'd usually find them in the performance-critical parts of APIs. Looking at some of the SQL these LLMs generate gives me that same feeling of dread. You spend a few hours just understanding it, then test it, then wonder whether to break it into smaller queries so you can actually feel confident shipping it to prod.