Google announced a major upgrade to Gemini 3 Deep Think, its specialized reasoning AI designed for scientific research and engineering applications. The updated system launched today for Google AI Ultra subscribers through the Gemini app, with API access available to select researchers and enterprises.
The company developed the enhanced reasoning mode in collaboration with scientists to address complex problems that lack clear solutions or involve incomplete data. Google states the system combines scientific knowledge with engineering applications to move beyond theoretical work toward practical uses.
Gemini 3 Deep Think achieved several benchmark results that redefine AI reasoning capabilities. It scored 48.4% on "Humanity's Last Exam" without external tools, a benchmark designed by experts to be nearly impossible for current AI.
The system also reached 84.6% on ARC-AGI-2, verified by the ARC Prize Foundation, which measures a model's ability to learn new skills and generalize to novel tasks.
In competitive programming, the model holds a 3455 Elo score on Codeforces, placing it in the "Legendary Grandmaster" tier reached by only a tiny fraction of human programmers globally. It achieved gold-medal level performance on the 2025 International Math Olympiad and demonstrated similar results on the written sections of the 2025 International Physics and Chemistry Olympiads.
The system scored 50.5% on the CMT-Benchmark for advanced theoretical physics, confirming deep knowledge in specialized scientific domains. Google's internal testing showed Gemini 3 Pro with 35% higher accuracy in resolving software engineering challenges than previous versions.
Real-world applications already demonstrate the system's practical utility. At Rutgers University, mathematician Lisa Carbone used Deep Think to review a technical mathematics paper focusing on Einstein's theory of gravity and quantum mechanics. The system identified a subtle logical flaw that had previously gone unnoticed during human peer review.
At Duke University's Wang Lab, researchers utilized the model to optimize fabrication methods for semiconductor materials. Deep Think successfully designed a precise recipe for growing thin films larger than 100 micrometers, a target that had previously eluded researchers using standard methodologies.
One of the most practical new features allows the model to analyze hand-drawn sketches and transform them into 3D-printable objects. By understanding geometry and physical requirements, Deep Think generates the necessary code to create functional files for 3D printing, streamlining the prototyping process for engineers.
Google's DeepMind division simultaneously introduced the AI agent Aletheia, designed specifically for mathematical research. A key feature of the agent is its ability to recognize when a problem cannot be solved, significantly saving researchers' time by avoiding fruitless exploration of impossible solutions.
DeepMind's subsidiary, Isomorphic Labs, also launched the IsoDDE engine for drug development. The system supports a wide range of complex molecules and offers a scalable framework for AI-driven drug design, building on the AlphaFold algorithm that predicted structures of over 200 million proteins in July 2022.
The updated Gemini 3 Deep Think represents a shift from general-purpose AI toward specialized tools capable of navigating the nuances of advanced academia. While standard models often struggle with "messy" data or problems lacking single clear-cut solutions, Deep Think is engineered to thrive in these gray areas using advanced "System 2" thinking.
Google is making the system available through the Gemini API to expand access for researchers and practitioners. This early access program allows enterprises, researchers, and independent engineers to integrate deep reasoning capabilities into custom applications.
The company emphasized that Deep Think is built not just for theoretical applications but also for practical engineering uses. The system helps researchers interpret complex datasets and enables engineers to model physical systems through code, targeting use cases that require advanced reasoning and structured data analysis.
Alphabet stock rose 1% Thursday following the announcement, bucking a broader tech selloff that saw the Nasdaq 100 fall 1.5%. The market response reflects investor confidence in Google's positioning within the competitive AI landscape against rivals like OpenAI and Anthropic.
