Google Releases Major Upgrade to Gemini 3 Deep Think Reasoning Mode

Google introduced a major upgrade to Gemini 3 Deep Think on February 12, 2026, expanding its capabilities to handle complex problems in science, research and engineering.

Credit: Google

The specialized reasoning mode is now available to Google AI Ultra subscribers through the Gemini app, where users select Deep Think in the prompt bar. Researchers, engineers and enterprises gained access to it through an early access program on the Gemini API.

The update follows Gemini 3's initial launch in November 2025 and builds on Deep Think's earlier performance, which included gold-medal results at the 2025 International Mathematics Olympiad and similar outcomes at the International Collegiate Programming Contest World Finals.

Google developed the upgrade in collaboration with scientists and research teams to address issues that involve incomplete or messy data and lack straightforward solutions.

Mathematician Lisa Carbone at Rutgers University applied the system to a highly technical mathematics paper and identified a subtle logical flaw that human peer reviewers had overlooked.

The Wang Lab at Duke University used it to refine fabrication processes for crystal growth and produced thin films larger than 100 micrometers, meeting a specific target that previous approaches had difficulty achieving.

Anupam Pathak, an R&D lead at Google’s Platforms and Devices division, tested the mode to speed up the design of physical components.

In one demonstration, Deep Think took a user sketch, analyzed its structure, generated a detailed 3D model and output a file ready for 3D printing.

The upgraded model recorded a score of 48.4% on Humanity’s Last Exam without external tools. It reached 84.6% on ARC-AGI-2, a result verified by the ARC Prize Foundation. It also achieved an Elo rating of 3455 on Codeforces, a platform for competitive programming problems.

DeepMind reported that Deep Think has contributed to 18 research problems spanning mathematics, physics and computer science. These include resolving bottlenecks in algorithms, settling a decade-old conjecture in online submodular optimization with a three-item counterexample, and deriving a novel solution for gravitational radiation from cosmic strings using Gegenbauer polynomials.

"Under direction from expert mathematicians and scientists, Gemini Deep Think is solving professional research problems across mathematics, physics, and computer science," the DeepMind team stated.

The system has also supported paper reviews for the STOC’26 computer science theory conference and generated an autonomous research paper on structure constants in arithmetic geometry.

Google directed interested parties to an early access form for the Gemini API at forms.gle/eEF5natXTQimPhYH9.