According to Dynamic Beating monitoring, the Google Research team released the text-to-SQL technology Gemini-SQL2 on June 12. The new system is built on the Gemini 3.1 Pro model and achieved an 80.04% execution accuracy on the industry-standard database query benchmark BIRD Single Model Leaderboard, setting a new performance record.
The earlier version Gemini-SQL, which was equipped with Gemini 2.5 Pro and utilized multi-task supervised fine-tuning, led the leaderboard with a 77.14% execution accuracy. The new version has improved the performance by 2.9 percentage points.
BIRD, as an industry benchmark for evaluating text-to-SQL system capabilities, consists of over 12,700 question-SQL pairs. The evaluation covers real-world database scenarios such as multi-table join queries, dirty data handling, and complex business context reasoning, with the metric of whether the generated SQL queries can execute correctly in a database.
Gemini-SQL2 was not released as a standalone large model but rather promoted as a foundational database query capability. Google plans to gradually integrate these capabilities into existing data services, including platforms such as BigQuery Studio, AlloyDB AI, and Cloud SQL Studio, to assist non-technical users in querying enterprise databases directly through natural language.
