Google’s Gemini SQL2 Tops the Charts at Turning Plain English Into Database Queries
Google Research has made a new tool. It is called Gemini SQL2. It just beat every rival at one tricky job. That job is turning plain English into database queries. A database is a big, organized store of information on a computer. A query is a request for some of that information. This skill is called text-to-SQL. It means you can ask a computer for data using normal words instead of code. On June 13, 2026, Google said the tool scored much higher than tools from OpenAI, Anthropic, and others on a popular test. The win was not close. It won by a lot.
What is text-to-SQL, in plain words?
Most company information sits inside databases. To pull answers out, people usually write SQL. SQL stands for Structured Query Language. It is a special kind of computer code used to ask databases questions. Writing it well takes training and lots of practice.
Text-to-SQL flips this around. You type a question in everyday words. For example, “Which products sold best last month?” The tool then writes the right SQL code for you. The database runs that code. Then it hands back the answer. The person asking does not need to know any code.
This sounds easy, but it is hard to do well. Google said that “turning natural language into correct SQL is especially hard because data is often layered.” Real databases have many tables. They have odd names and hidden rules. One small mistake can give a wrong answer. Or it can stop working completely.
How well did Gemini SQL2 do?
Google tested Gemini SQL2 on a test called BIRD. BIRD is a benchmark. A benchmark is a shared test. It lets people compare different AI tools in a fair way. BIRD checks how often a tool’s SQL both looks right and really runs to give the correct answer. This is called execution accuracy. It just means how often the code works and gives the right result.
Gemini SQL2 scored 80.04% and came in first place. The next best tool was OpenAI’s GPT-5.5-xhigh. It scored 72.8%. Anthropic’s Claude Opus 4.6 scored 70.9%. Tools from Databricks, AWS, Tencent, and Alibaba came even further behind. So Gemini SQL2 beat the second-best tool by more than seven points.
Key facts at a glance
| Detail | What we know |
|---|---|
| Tool name | Gemini SQL2 |
| Built by | Google Research |
| Announced | June 13, 2026 |
| Built on | Gemini 3.1 Pro |
| Job | Text-to-SQL (plain English to database queries) |
| Benchmark used | BIRD (execution accuracy) |
| Top score | 80.04% |
| Public release | None announced yet |
Benchmarks and specs: how the models compare
Here are the BIRD scores for the top tools. These are the exact numbers Google shared. We have not added any made-up numbers.
| Model | Maker | BIRD execution accuracy | Rank |
|---|---|---|---|
| Gemini SQL2 | Google Research | 80.04% | 1st |
| GPT-5.5-xhigh | OpenAI | 72.8% | 2nd |
| Claude Opus 4.6 | Anthropic | 70.9% | 3rd |
| Other models | Databricks, AWS, Tencent, Alibaba | Below the top three (exact scores not reported) | Lower |
What it means: On this one test, Gemini SQL2 is clearly ahead. It writes correct, working database queries from plain English more often than its rivals do.
How Gemini SQL2 works
Gemini SQL2 is built on Gemini 3.1 Pro. That is one of Google’s main AI tools. An AI model is a computer program that learns from lots of data. The tool is set up to handle messy, real-world databases. That means lots of tables, tricky links, and business rules baked into the data.
Google says the goal is queries where “the generated SQL queries both look correct and execute successfully.” In short, the code should not just look right. It must really run and give back the right answer.
What Google has not shared yet
You cannot use Gemini SQL2 right now. There is no public release. That means it is not open for everyone to use yet. You cannot download it or try it today. Google has also not put out a research paper. A research paper is a detailed report that explains how something works. So we know the score, but not all the steps behind it.
Google did give a hint about the future. Better SQL skills could power new “ask in plain words” features across its data tools. That could let more people ask data questions using normal words.
FAQ
What is Gemini SQL2?
It is a text-to-SQL tool from Google Research. It turns questions written in plain English into working database queries. It is built on the Gemini 3.1 Pro model.
What score did it get?
It scored 80.04% execution accuracy on the BIRD test. That put it in first place. The next best tool, OpenAI’s GPT-5.5-xhigh, scored 72.8%.
Can I use Gemini SQL2 now?
Not yet. Google has not announced a public release. There is also no research paper out about it so far.
Why is text-to-SQL hard?
Real databases are layered and complex. They have many tables, strange names, and hidden business rules. One small slip can give a wrong answer or fail to run.
Why it matters (especially for India and founders)
In many Indian startups, the data sits in databases. But only a few people can read it. Founders, sales teams, and students often have to wait. They ask engineers to pull simple reports for them. Better text-to-SQL could end that wait.
Imagine a small shop owner asking, “How many orders came from Delhi last week?” And getting the answer right away. No engineer needed. That saves time and money. This matters a lot for small teams.
It also makes data easier to use. Students and new staff could explore data without years of SQL training. As tools like this get better, data could become something everyone in a company can use. Not just coders.
The takeaway
Gemini SQL2 shows that AI is getting much better at one useful job. That job is writing correct database queries from plain English. The 80.04% BIRD score is a clear lead today. But it is not open to the public yet. So the real test will be how it works once people outside Google can use it.
Source: The Decoder
