Google Rolls Out Deep Think AI to Gemini Ultra Subscribers

Details

• Google has launched Deep Think, an enhanced reasoning mode built on Gemini 2.5 Pro, for AI Ultra subscribers through the Gemini app, targeting advanced problem-solving such as mathematical discovery and complex coding.
• The rollout involves collaboration with mathematicians, who are receiving the gold-medal-winning IMO model to provide expert feedback and drive further enhancements.
• Deep Think utilizes parallel thinking, allowing it to explore multiple solution paths simultaneously, and benefits from extended reasoning times and reinforcement learning processes.
• This new mode offers notably faster, real-time performance but achieves bronze-level accuracy, in contrast to Google's gold-medal IMO model, which is optimized for peak benchmark results with longer inference times; it also represents a major upgrade over the earlier AlphaGeometry and AlphaProof systems announced in 2024.
• The AI's advances were demonstrated through top performance on coding and academic benchmarks like LiveCodeBench V6 and Humanity’s Last Exam, showcasing strengths in algorithm design and literature analysis.

Impact

The launch of Deep Think highlights Google's aggressive push in advanced AI reasoning, setting the stage for direct competition with OpenAI’s own IMO-level breakthroughs. By combining rapid, parallel reasoning with close academic partnerships, Google is signaling both commercial ambition and a commitment to cutting-edge research. This dual-track approach could shape future expectations for AI applications in science and enterprise domains.