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

Google has announced a major upgrade to its specialised reasoning model, Gemini 3 Deep Think, claiming the system can now address complex real-world problems across science, research, and engineering.

Announcing the development on X (formerly Twitter), Google CEO Sundar Pichai said the enhanced model is designed to handle “tough, real-world challenges.”

The company said it worked closely with scientists and researchers while developing the upgraded system to ensure it can contribute to actual scientific and technical work rather than only theoretical tasks.

According to Google, the model combines deep domain knowledge with applied engineering capabilities. The company said the system “blends deep scientific understanding with practical engineering skills to solve actual problems and support real-world applications.”

Strong performance across benchmarks

Google claimed that Gemini 3 Deep Think posted leading scores across several demanding academic and technical benchmarks.

Google said the model has achieved “gold-medal standards” in mathematics and programming competitions. It scored 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.

The company added that the model’s strengths extend beyond narrow domains.

“Beyond mathematics and competitive coding, Gemini 3 Deep Think now also excels across broad scientific domains such as chemistry and physics,” Google said.

The system also achieved gold-medal-level results at the International Physics Olympiad 2025 and the International Chemistry Olympiad 2025, and scored 50.5% on the CMT Benchmark, which evaluates performance on complex theoretical physics problems. 

Real-world use cases

Google said that the model is designed for practical use, including analysing complex datasets and building simulations of real-world systems.

Lisa Carbone, a mathematician at Rutgers University, used Gemini 3 Deep Think to analyse a technical research paper exploring the intersection of Einstein’s theory of gravity and quantum mechanics. During testing, the model reportedly identified logical flaws that had been missed in human review.

Separately, researchers at the Wang Lab at Duke University used the system to optimise the growth of specialised semiconductor crystals. The model generated a detailed process for producing thin films larger than 100 micrometres.

Availability

Google said Gemini 3 Deep Think will initially be available to Google AI Ultra subscribers and to developers through the Gemini API via an early access programme.