Google's Gemini-SQL2 Dominates Text-to-SQL Benchmarks with Record 80% Accuracy

Summary

Google Research unveiled Gemini-SQL2, a new text-to-SQL system built on Gemini 3.1 Pro that translates natural language into executable SQL database queries. The model achieved an execution accuracy of 80.04 percent on the BIRD benchmark, significantly outperforming competitors—OpenAI's GPT-5.5-xhigh scored 72.8 percent and Anthropic's Claude Opus 4.6 reached 70.9 percent. Models from Databricks, AWS, Tencent, and Alibaba all performed considerably worse.

The task of converting natural language to correct SQL is exceptionally challenging because real-world data often has complex layered structures and queries must account for intricate business logic. Google's achievement represents a meaningful advance in this difficult domain, with the company noting that the generated SQL queries both appear correct and execute successfully. The research team indicated that improved SQL understanding could enhance natural language features across Google's data services more broadly. However, no public release of the model or peer-reviewed research paper has been announced, limiting immediate accessibility for developers and researchers.

Editorial Opinion

This benchmark result demonstrates Google's continued strength in LLM-driven reasoning tasks, particularly for structured data translation. However, without a public release or published paper, the practical impact remains uncertain—the AI research community will be watching to see if Google makes the model or detailed methodology available. If Gemini-SQL2 eventually becomes public, it could set a new standard for text-to-SQL systems and raise competitive pressure across the industry.