Google introduced an upgraded version of Gemini 3 Deep Think, its advanced reasoning mode aimed at tackling complex scientific and engineering problems, and said it would broaden access to the system through its developer platform.
The company said the new iteration was refined in collaboration with scientists and academic researchers to better address open-ended research tasks where data may be incomplete and solutions are not clearly defined. “We updated Gemini 3 Deep Think in close partnership with scientists and researchers to tackle tough research challenges — where problems often lack clear guardrails or a single correct solution and data is often messy or incomplete,” Google said in a statement.
See also: Google Rolls Out Gemini AI Features in Chrome to Boost Web Multitasking
Gemini 3 Deep Think builds on prior versions that demonstrated strong results in mathematics and programming competitions. Google said the latest model set “a new standard (48.4%, without tools) on Humanity’s Last Exam,” achieved 84.6% on ARC-AGI-2 as verified by the ARC Prize Foundation, and reached an Elo rating of 3455 on Codeforces. The company added that it attained gold-medal-level performance at the 2025 International Math Olympiad.
Beyond competitive benchmarks, Google said the system has improved capabilities in scientific disciplines including physics and chemistry. According to the company, Deep Think demonstrated gold medal-level results on written sections of the 2025 International Physics and Chemistry Olympiads and scored 50.5% on the CMT-Benchmark for theoretical physics.
See also: Amazon Considers Content Licensing Marketplace for AI Firms
The upgraded reasoning mode is now available in the Gemini app for Google AI Ultra subscribers. Google also said it is, for the first time, offering access to select researchers, engineers and enterprises through the Gemini API under an early access program, as it seeks to accelerate real-world applications of advanced AI systems.
Source: Google