Google Upgrades Gemini 3 Deep Think for Science and Engineering Applications

Details

• Google announced a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode, blending deep scientific knowledge with engineering utility to drive practical applications beyond abstract theory.
• Developed in partnership with scientists and researchers to tackle challenges lacking clear solutions or with messy data, such as interpreting complex data and modeling physical systems.
• Achieves top benchmarks: 48.4% on Humanity’s Last Exam (without tools), 84.6% on ARC-AGI-2 (verified by ARC Prize Foundation), Elo 3455 on Codeforces, gold-medal performance on International Math Olympiad 2025, and strong results in physics and chemistry olympiads.
• Practical example: Converts a hand-drawn sketch into a 3D-printable model by analyzing the shape and generating a file.
• Now available in the Gemini app for Google AI Ultra subscribers; also offered via Gemini API early access for researchers, engineers, and enterprises.

Impact

Google's Gemini 3 Deep Think upgrade positions it as a leader in AI-driven scientific research, setting new benchmarks like 84.6% on ARC-AGI-2 that surpass prior frontier models and pressure rivals such as OpenAI's o1 and Anthropic's Claude, which have targeted similar reasoning feats but lag in verified AGI-like scores. By enabling practical tools like sketch-to-3D printing and API access for enterprises, it lowers barriers for real-world R&D, accelerating discovery in messy domains like theoretical physics where data scarcity persists. This aligns with trends in agentic AI and on-device reasoning, potentially shifting market dynamics toward hybrid science-engineering models and steering funding toward multimodal benchmarks over the next 12-24 months as competitors race to match its gold-medal olympiad performance.