Google AI Unveils Gemini 3 Deep Think with Record-Breaking Reasoning (GOOGL)

Google (GOOGL) on February 12, 2026, announced a major upgrade to Gemini 3 Deep Think, a specialized reasoning model, now achieving superhuman performance in complex scientific and engineering tasks. The model set new benchmarks: it reached an Elo score of 3,455 on Codeforces, surpassed human experts on the Humanity's Last Exam (HLE) with 48.4%, and scored 84.6% on ARC-AGI-2. In practical research, Deep Think identified a subtle logical flaw in a high-energy physics paper review and enabled breakthroughs in materials science and physics modeling, including faster 3D printable prototypes from hand sketches. The update was developed with contributions from former Anthropic researchers, and early access is available through the Gemini API to researchers and engineers.