Google upgrades Gemini 3 Deep Think to tackle real-world scientific problems, hits top benchmarks

Google has unveiled a significant upgrade to its specialized reasoning model, Gemini 3 Deep Think, asserting that it is now equipped to tackle intricate real-world issues in the fields of science, research, and engineering. In a recent announcement on X (formerly Twitter), CEO Sundar Pichai emphasized that this enhanced model is crafted to address "tough, real-world challenges." The development of Gemini 3 Deep Think involved close collaboration with scientists and researchers to ensure it offers tangible contributions to scientific and technical endeavors, moving beyond purely theoretical applications. Google highlights that this model intricately merges deep domain knowledge with practical engineering capabilities, stating that it blends scientific understanding with the skill set needed for real-world problem-solving. In terms of performance, Google claims that Gemini 3 Deep Think has achieved remarkable scores on various rigorous academic and technical benchmarks. The model reportedly reached "gold-medal standards" in both mathematics and programming contests, scoring an impressive 48.4% on Humanity’s Last Exam, 84.6% on ARC-AGI-2, and an Elo rating of 3455 on the competitive coding platform Codeforces. Furthermore, its talents extend into broader scientific areas such as chemistry and physics, achieving gold-level results in the International Physics Olympiad 2025 and the International Chemistry Olympiad 2025, along with a score of 50.5% on the CMT Benchmark, which assesses performance on complex theoretical physics problems. Google envisions practical applications for this model, including the analysis of intricate datasets and the creation of simulations for real-world systems. For instance, mathematician Lisa Carbone from Rutgers University employed Gemini 3 Deep Think to evaluate a technical research paper focused on the convergence of Einstein’s theory of gravity and quantum mechanics. Remarkably, the model was able to pinpoint logical inconsistencies that had eluded human scrutiny. Additionally, researchers at the Wang Lab at Duke University utilized the system to enhance the growth of specialized semiconductor crystals, developing an intricate process for producing thin films exceeding 100 micrometres. Initially, Gemini 3 Deep Think will be accessible to Google AI Ultra subscribers and developers through an early access program via the Gemini API, marking a significant step forward in practical AI applications.