SHELFCRITTER AI Digest

MODEL

Gemini SQL2

modeltopic-notegoogletext-to-sql

Overview

Gemini SQL2 is Google Research’s June 2026 text-to-SQL specialisation built on top of Gemini 3 Pro (3.1 Pro variant) via prompting and scaffolding with no fine-tuning. The system reaches 80.04% execution accuracy on the BIRD single-model leaderboard — the first single-model number above 80% on the canonical text-to-SQL evaluation, ~7 points clear of GPT-5.5-xhigh at ~72.8% and Claude Opus 4.6 at ~70.9%, with the prior best at 77.2%. The structural read is that the win comes from prompting on top of an existing flagship, not a new pre-trained checkpoint.

Timeline

  • 2026-06-14-AI-Digest — Google Research announces Gemini SQL2 on 2026-06-12. Reaches 80.04% on BIRD execution accuracy vs GPT-5.5-xhigh ~72.8% and Claude Opus 4.6 ~70.9% (prior best 77.2%). BIRD (Big Bench for Large-scale Database Grounded Text-to-SQL) is the canonical text-to-SQL eval; the ~7-point gap over the next-best general-purpose model is the headline, and the more interesting structural read is that the win comes from prompting / scaffolding on top of an existing flagship (Gemini 3 Pro 3.1 Pro variant), not a new pre-trained checkpoint. Disciplined caveat: BIRD execution accuracy is not BIRD-Pro or BIRD-Critic — the harder slices (multi-table joins, long schemas) tend to compress these gaps.

Key Developments

  1. First Single-Model BIRD Result Above 80%: 80.04% execution accuracy on BIRD is the first single-model crossing of 80% on the canonical text-to-SQL leaderboard, ~7 points clear of GPT-5.5-xhigh. Text-to-SQL is the enterprise-data integration surface most directly tied to coding-agent revenue (every agent ultimately needs to talk to the company’s database), so a single-model leaderboard win above 80% materially compresses the practical gap between agent-mediated SQL generation and a human analyst writing the query.

  2. Prompting/Scaffolding on Top of Gemini 3 Pro, Not a New Checkpoint: The structural piece worth pinning is that Gemini SQL2 is built on Gemini 3 Pro (3.1 Pro variant) without fine-tuning — the win is in the scaffolding, not the base model. This is a different shape of result than a new pre-trained model topping the leaderboard, and it’s the kind of result that other labs can chase on their own flagships without retraining.

  3. A Third, Narrower Leaderboard Where Google Wins: Lands alongside the frozen Aider polyglot top-5 and the SWE-Bench Verified top-3 (both stuck on closed-reasoning models that are currently disabled or unchanged) as a third, narrower leaderboard now in the mix — with a Google win where the general-purpose leaderboards split.

See also: Google, Gemini 3 Pro, GPT-5, Claude Opus 4.8, MOC - Developer Tools, MOC - Major Companies.