Google announces Gemini 3 Deep Think: an upgraded deep thinking large model advancing scientific, research, and engineering applications, capable of reaching gold medal levels in mathematics, physics, and chemistry competitions

K-LinePoet · 2026-03-05T21:57:03+00:00

Google has made significant upgrades to Gemini 3 Deep Think, introducing a "Reasoning Mode" specifically designed for research and engineering scenarios. The new version enhances capabilities in fields such as mathematics, programming, chemistry, and physics, and is available to Google AI Ultra users starting February 12. It is also accessible to some researchers through the Gemini API for early testing. Google aims to promote practical research applications through this mode, helping to handle complex data and modeling.

K-LinePoet

2026-03-05 21:57:03

Abstract generation in progress

Tech Home, February 13 — Google announced a major upgrade to Gemini 3 Deep Think last night, claiming it is a “reasoning mode” specifically developed for scientific, research, and engineering scenarios, aimed at advancing the frontiers of intelligence.

According to the company, the new version of Deep Think was developed collaboratively by Google engineers and scientists from various industries. Its goal is to address common complex problems in real research environments: lacking clear boundaries, not necessarily having a single solution, and often dealing with messy or incomplete data.

A key change in this upgrade is the expanded availability of Deep Think. Google states that the updated Deep Think has been available since February 12 local time in the Gemini app to Google AI Ultra subscribers.

Tech Home notes that Google is initially providing Deep Think to some researchers, engineers, and enterprises via the Gemini API through an “early access program,” with an application portal open to recruit interested users.

In terms of performance, Google emphasizes that the new Deep Think continues to improve in high-level reasoning tasks such as mathematics, algorithms, and programming. The model scored 48.4% on Humanity’s Last Exam without using tools; achieved 84.6% on ARC-AGI-2, verified by the ARC Prize Foundation; earned an Elo rating of 3455 on Codeforces programming benchmarks; and is expected to reach gold medal level at the 2025 International Mathematical Olympiad (IMO 2025).

Beyond mathematics and programming, Google also highlights that the new Deep Think demonstrates stronger capabilities in scientific fields like chemistry and physics. The company states that the model achieved gold medal results in the written portions of the 2025 International Physics and Chemistry Olympiads, and scored 50.5% on the CMT-Benchmark related to theoretical physics.

Google says that Deep Think’s goal is not only to excel in benchmark tests but also to promote practical engineering and research applications, such as helping researchers interpret complex data and assisting engineers in modeling physical systems through code. Google plans to continue bringing Deep Think to researchers and practitioners “where it is most needed,” and considers opening early access via the Gemini API an important step.

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.