I’ve always hated SQL, but fortunately LLMs write it so well that it’s effectively become a read-only language now. You just need to know enough to check the output.
I agree. Claude Code writes superb SQL queries for very complex data. I was dealing with PostgreSQL recently, and it improved the query from 30 seconds to 5 seconds. I couldn't figure it out myself.
Prompting with documentation and examples works. In an agentic tool having an MCP server for the db helps assuming it is a straightforward schema with explicitly defined relationships. Also helps if the tables correspond to entities in a natural way.
We're pretty locked down around AI tools. Right now we can only really use GH Copilot, it's been ok so far though it's funny to see it suggest edits then suggest the opposite on the next time it review the PR if you accept them.
sonnet 4.5 was really bad at anything more than simple queries. even GPT 5 was not great. gemini was consistently good even at 2.5; caught multiple bugs in outputs of either. I haven't tested Opus 4.5 properly at SQL yet, but I've got a feeling Anthropic doesn't prioritize it in training and google does.
I switched to Opus permanently, but I am being cautious. It tends to go berserk and add a bunch of extras. Most of my time is spent narrowing the task to it. Sonnet was much more straightforward forward, but Opus' quality is unmatched when compared. I caught Opus messing up my CSP settings, and I had to restrict the access levels to deeper levels. Using it as an agent with full access without a plan more is suicide with Opus, but Sonnet didn't have this problem.
