Google Enhances Gemini Deep Think, Launches AI Mathematician, and Accelerates Drug Design

Google has unveiled significant enhancements to its Gemini 3 Deep Think model, aimed at addressing complex scientific and engineering challenges. In comparative tests, the updated model has shown superior performance against competitors such as OpenAI's GPT-5.2 and Anthropic's Claude Opus 4.6, particularly in various assessments including visual puzzles and theoretical physics benchmarks.

According to Google's blog, the updates were made in collaboration with scientists to tackle intricate problems that often lack definitive solutions. The Gemini 3 Deep Think excels in mathematics and programming, demonstrating skills comparable to top international competition gold medalists. In the CMT-Benchmark, it achieved a score of 50.5%, highlighting its strong grasp of theoretical physics.

"Deep Think is not only advanced in performance but also practical, assisting researchers in interpreting complex data and enabling engineers to model physical systems," Google stated.

The updated version is accessible through the Gemini app for Google AI Ultra subscribers and via the Gemini API for select developers.

Introduction of AI Mathematician Aletheia

In addition, Google’s DeepMind has launched a new AI agent named Aletheia, which has achieved a record-setting success rate of 91.9% on the IMO-ProofBench Advanced benchmark, a challenging mathematics test. Built on the Gemini Deep Think framework, Aletheia includes a verification module that helps refine solutions by identifying errors and suggesting improvements.

Aletheia is designed to recognize unsolvable problems, thereby saving valuable time for researchers. By utilizing Google Search, it enhances the accuracy of its scientific references, reducing errors in computational processes.

Notable accomplishments of Aletheia include:

• the complete generation of a scientific paper on structural constants in arithmetic geometry;
• collaborative proof for estimates involving systems of interacting particles;
• independent solutions to four problems from the Erdős list, including one previously unsolved.

DeepMind asserts that Aletheia’s achievements underscore the efficacy of scaling laws in proof-based mathematics.

Advancement in Drug Design Technology

Moreover, DeepMind’s Isomorphic Labs has launched the IsoDDE engine for drug design, which has demonstrated twice the prediction accuracy of AlphaFold 3 in complex tests. While AlphaFold made significant strides in predicting protein structures, IsoDDE marks a new frontier in drug design.

Key features of IsoDDE include:

• enhanced accuracy in predicting binding affinity compared to traditional methods;
• the ability to identify concealed structures within proteins for potential drug binding;
• support for a diverse array of complex molecules, including large biological structures and antibodies.

"IsoDDE provides a scalable foundation for AI-driven drug design, ensuring the prediction accuracy required for new biological systems with unmatched reliability," the company noted.

Previously, AlphaFold had predicted the structures of over 200 million proteins, encompassing nearly all known compounds found in various organisms.