Google announced a significant evolution of its specialized reasoning model, Gemini 3 Deep Think, introducing expanded capabilities designed to tackle some of the most complex challenges in modern science, research and engineering. This updated mode is now available to Google AI Ultra subscribers in the Gemini app, and for the first time is being offered through an early access program via the Gemini API for select researchers, engineers and enterprises.
Developed in close collaboration with scientists and research professionals, the enhanced Deep Think mode is engineered to handle problems that lack clear guardrails and often involve incomplete or messy data, moving beyond theoretical tasks to deliver practical, real-world outcomes.
Breakthrough Reasoning and Scientific Problem Solving
The upgraded Deep Think builds on its predecessor’s capabilities by blending deep scientific understanding with applied engineering utility. It demonstrates improved reasoning performance across rigorous academic and industry-standard benchmarks, including:
- A new record 48.4% (without tools) on Humanity’s Last Exam a benchmark designed to test the limits of modern frontier models
- 84.6% on ARC-AGI-2, verified by the ARC Prize Foundation a major general reasoning milestone
- 3455 Elo on Codeforces competitive programming challenges
- Gold-medal performance on the International Math Olympiad 2025
In addition to mathematical and coding excellence, the model now shows superior performance in broad scientific domains such as physics and chemistry including gold-medal-level results on the written segments of both the 2025 International Physics and Chemistry Olympiads, as well as advanced theoretical physics proficiency on the CMT-Benchmark.
Also Read: Tines Unveils AI Interaction Layer to Operationalize Enterprise AI at Scale
Real-World Engineering and Innovation
This emphasis on practical applications has been reflected in the latest version of the Gemini 3 Deep Think tool, which is intended for helping engineers and scientists solve complex problems such as interpreting complex data sets, coding a physical system, and turning concept designs into 3-D printable objects.
Use cases from early testers highlight Deep Think’s impact:
- At Rutgers University, mathematician Lisa Carbone used Deep Think to uncover subtle logical flaws in advanced technical work previously missed in human peer review.
- At Duke University’s Wang Lab, the model optimized complex crystal-growth fabrication methods, designing processes that met precise size targets for semiconductor research.
- Anupam Pathak, R&D lead at Google’s Platforms and Devices division, applied Deep Think to accelerate design iteration for physical component engineering.
Expanded Availability and Early Access
Starting now, the improved Deep Think mode is integrated into the Gemini app for all Google AI Ultra subscribers. In addition, scientists, engineers and enterprise teams can now express interest in early access to test Deep Think through the Gemini API opening opportunities for deeper integration into research workflows and enterprise systems.
