Gemini-SQL2: Natural Language Becomes Executable SQL Queries

What changed: Google Research redefines text-to-SQL

On June 13, 2026, Google Research announced Gemini-SQL2, an advanced text-to-SQL conversion system. The model is built on Gemini 3.1 Pro and represents a qualitative leap over previous generations. According to The Decoder, Gemini-SQL2 reached 80.04% accuracy on the BIRD benchmark, the most widely used in the industry for evaluating text-to-SQL systems.

This result consistently beats competing models from OpenAI and Anthropic. Google is therefore positioning itself as the technical leader in this specific area of AI applied to data. The BIRD benchmark measures a system's ability to generate correct SQL queries from natural-language questions on real, complex databases.

Google has also stated that the technology behind Gemini-SQL2 will be integrated into its existing data services, including BigQuery, Looker and other tools in the Google Cloud suite. As a result, companies already using the Google ecosystem could benefit from these capabilities without adopting new tools.

The BIRD benchmark and what 80% accuracy means

The BIRD benchmark (Big Bench for Large-scale Database Grounded Text-to-SQL Evaluation) is the industry's reference standard. It evaluates models on real databases, with complex schemas and ambiguous questions. Reaching 80% on this test is no small feat.

Previous systems scored noticeably lower. According to the published data, the gap between Gemini-SQL2 and its competitors is several percentage points. Still, it's important to put this in context: the remaining 20% of errors can still produce incorrect or incomplete queries in production environments. Human oversight therefore remains necessary in critical scenarios.

To learn more about how AI benchmarks work technically, see the analyses published by MIT Technology Review, which has covered the topic of evaluating language models on structured tasks several times.

Immediate impact for Italian SMEs without a technical team

For many Italian B2B and retail SMEs, access to company data still goes through technical staff. A sales manager who wants to know which customers bought twice in the last quarter has to wait for a developer to write the query. That slows down decisions.

Gemini-SQL2 could change this dynamic. A system that turns questions in Italian — or English — directly into executable SQL significantly lowers the technical barrier. Moreover, the 80% accuracy on the BIRD benchmark suggests the system works well even on databases with non-trivial structures.

Likewise, tools like these fit with the AI applied to business trends that we at SHM Studio follow for our clients. In particular, automating data access is one of the most concrete, high-return use cases for mid-sized companies. That's why it's worth evaluating how this technology fits into existing operational workflows.

How it fits into the Google Cloud ecosystem

Google has indicated that Gemini-SQL2 will be rolled out within its cloud services. BigQuery, Google Cloud's data warehouse, is the most obvious candidate for a first integration. Looker, the business intelligence platform acquired by Google, could benefit even more directly.

As a result, SMEs that have already invested in the Google ecosystem could access these features through gradual updates to existing services. However, commercial release timing hasn't been officially announced yet. It's therefore premature to plan operational integrations in the short term without further announcements.

If you manage campaigns and data on Google platforms, it's also worth watching developments around Google Ads and its automated reporting features, which could indirectly benefit from this technology. Likewise, if you work with data-driven digital marketing, you'll want to follow how these tools evolve.

The competitive landscape: OpenAI and Anthropic lag behind on this specific task

Gemini-SQL2's result is particularly significant because it comes at a time when OpenAI and Anthropic dominate public perception of the AI sector. Yet the BIRD benchmark tells a different story in this specific domain.

OpenAI's models — including GPT-4o and recent versions — and Anthropic's, such as Claude 3.7, haven't reached comparable levels on structured text-to-SQL. Google thus shows that vertical specialization on a specific task can produce measurable competitive advantages. That's a meaningful strategic signal for the market.

According to Gartner's analyses of technology trends, specializing AI models for vertical domains is one of the most promising directions for 2026-2027. Gemini-SQL2 is part of a broader trajectory of technical differentiation among the major players.

What to do now: three practical considerations for SMEs

First, map your existing company databases and identify which queries are most often requested from technical staff. This exercise lets you estimate the potential time savings a text-to-SQL tool could generate.

Next, assess whether your company's data infrastructure is already on Google Cloud or whether migration plans exist. Integrating Gemini-SQL2 will be smoother for those already in the Google ecosystem. If you use on-premise databases or other cloud providers, accessing this technology may require intermediate steps.

Finally, we recommend not waiting for the commercial release to start structuring your data to make it more accessible. A solid data architecture is a prerequisite for any natural-language query tool. At SHM Studio, we can support SMEs in this assessment and planning phase, as part of our AI consulting and data-driven digital marketing services.

What the press releases don't say: real limits to consider

Benchmarks are useful, but they don't tell the whole story. The 80.04% accuracy on BIRD is a result in a controlled environment. In real business settings, databases have non-standard column names, implicit relationships and undocumented internal conventions. Real-world performance may therefore be lower than what's measured in the lab.

Moreover, the quality of the generated queries depends heavily on the quality of the question asked. A non-technical user might phrase ambiguous requests, producing imprecise results even with an excellent model. Training end users therefore remains a critical factor in the success of these tools.

Even so, the direction is clear. The gap between natural language and structured data is closing fast. For Italian SMEs, this means that investing today in the quality and organization of your data, through services like data-driven web development or SEO strategies built on structured analytics, is a strategic choice with an ever shorter payback horizon.

Outlook: where this technology will lead in 2027-2028

In the short term, the most likely development is the integration of Gemini-SQL2 into Google Cloud products. In the medium term, similar features are expected to become standard in the main business intelligence platforms, from Tableau to Power BI.

As a result, by 2027-2028, querying company databases in natural language could become a basic feature rather than a differentiator. Consequently, the competitive advantage will shift from being able to access data to being able to interpret it and act on it quickly.

To dig deeper into AI applied to Italian SMEs, there's a useful read on the SHM Studio blog, where we regularly publish analysis on these topics. If you want to discuss how to integrate AI tools into your business processes, you can contact us directly. And if you manage content and digital communication, our SEO copywriting and LinkedIn campaigns services can help you tell the story of these changes to customers and stakeholders.