Google Research has put a number on database querying that no one had hit before. Gemini-SQL2, announced June 12, is a text-to-SQL system built on Gemini 3.1 Pro that turns plain-language questions into execution-ready SQL — and it is the first to clear 80% on the BIRD single-model leaderboard, landing at 80.04%.
## What Gemini-SQL2 does
The job is narrow but valuable: take “which customers spent the most last quarter” and produce SQL that actually runs against a real schema. BIRD is the hard benchmark for this because it uses large, messy, real-world databases rather than toy tables, so an 80% execution-accuracy score is a meaningful jump for anyone who wants non-engineers talking to their data directly.
## The catch
It is a milestone, not a download. Google has not said when or where you can use Gemini-SQL2, or whether it ships outside Google’s own products at all. Commentators have also flagged the obvious enterprise risk: a model that writes runnable SQL against production data needs tight guardrails before it touches anything live.
Top AI Product 
Leave a comment