Google has unveiled a major update to its Gemini 3 Deep Think mode, designed to address complex research challenges in science, engineering, and academic domains. The mode integrates deep scientific knowledge with practical engineering capabilities to analyze incomplete or messy data and identify solutions that extend beyond standard AI reasoning.
The update was developed in partnership with scientists and researchers, enabling Deep Think to tackle problems lacking clear guardrails or single correct answers. The tool is now available within the Gemini app for Google AI Ultra subscribers and, for the first time, via the Gemini API for select researchers, engineers, and enterprises. Early access is available through an application process.
Early Use Cases Highlight Research Impact
Early testers report that Deep Think can analyze complex, technical content and identify subtle errors previously undetected in human peer review. Lisa Carbone, a mathematician at Rutgers University, used Deep Think to review a paper in high-energy physics bridging general relativity and quantum mechanics, and the model successfully highlighted a logical flaw overlooked by prior reviewers.
The mode also demonstrates advanced reasoning capabilities in mathematics, coding, and competitive benchmarks. Key performance highlights include:
- 48.4% on Humanity’s Last Exam without tools, a frontier benchmark
- 84.6% on ARC-AGI-2, verified by the ARC Prize Foundation
- Elo rating of 3455 on Codeforces competitive programming challenges
- Gold-medal level at the 2025 International Math Olympiad
Expanding to Scientific Domains and Engineering
Beyond math and coding, Gemini 3 Deep Think performs at high levels across chemistry and physics. It achieved gold-medal performance on the written sections of the 2025 International Physics Olympiad and Chemistry Olympiad and scored 50.5% on the CMT-Benchmark in advanced theoretical physics.
The update also emphasizes real-world engineering applications. Deep Think can interpret sketches and generate 3D-printable models, enabling practical implementation of complex designs. Researchers can now leverage the mode for modeling physical systems, analyzing experimental data, and automating intricate engineering tasks.
Availability and Access
The updated Deep Think mode is immediately available to Google AI Ultra subscribers in the Gemini app. Select researchers, engineers, and enterprises can apply for early access via the Gemini API. Google said the release marks the beginning of broader availability for scientific and engineering applications, signaling a push to bring AI-driven reasoning tools to professional workflows at scale.
The launch also comes amid growing competition in AI research workspaces, following OpenAI’s introduction of Prism, a free AI-powered scientific platform integrated with GPT-5.2 that assists researchers with writing, reviewing, and visualizing scientific papers.
