From reasoning models that win math olympiads to quantum algorithms approaching real-world use, Google just published its most comprehensive year-end research review yet.
The tech giant’s leadership team, including Chief Scientist Jeff Dean, DeepMind CEO Demis Hassabis, and SVP James Manyika, outlined eight areas where the company made significant progress. The throughline is clear: 2025 marked the shift from AI as a tool to AI as a utility that thinks, acts, and explores alongside humans.
Gemini 3 Sets New Standards
The headline achievement centers on the Gemini model family. Gemini 3 Pro topped the LMArena Leaderboard and posted breakthrough scores on Humanity’s Last Exam, a benchmark designed to test whether AI can genuinely reason like humans.
The model achieved 23.4% on MathArena Apex, setting a new state-of-the-art for frontier models in mathematics.
Gemini 3 Flash followed in December, combining Pro-grade reasoning with dramatically lower latency and cost. Google emphasized a pattern: each generation’s Flash model now surpasses the previous generation’s Pro model.
The open-source Gemma family also received major upgrades, with Gemma 3 becoming the most capable model runnable on a single GPU or TPU.
Products Gain Agentic Capabilities
Google transformed its product portfolio with AI features throughout 2025. The Pixel 10 shipped with nine AI-powered capabilities. Search introduced AI Mode. NotebookLM added Deep Research functionality.
The company also launched Google Antigravity, marking what it calls “a new era in AI-assisted software development.” The platform moves beyond coding assistance toward agentic systems that collaborate with developers directly.
Scientific Milestones Stack Up
AlphaFold celebrated its fifth anniversary with striking adoption numbers: over 3 million researchers in more than 190 countries have used the Nobel-winning protein structure prediction system. More than 1 million users come from low- and middle-income countries.
New systems joined the portfolio. AlphaGenome launched in June for genomic analysis. DeepSomatic arrived in October for identifying genetic variants in tumors. An AI co-scientist tool debuted in February to accelerate scientific breakthroughs.
Gemini’s Deep Think capability achieved gold-medal standard at both the International Mathematical Olympiad and the International Collegiate Programming Contest World Finals.
Quantum Computing Progress
The quantum research division drew unusual public attention this year. The Quantum Echoes algorithm represented a major step toward practical quantum computing applications.
Googler Michel Devolet received the 2025 Physics Nobel Prize alongside former Googler John Martinis and UC Berkeley’s John Clarke for their foundational quantum research from the 1980s.
Hardware and Robotics Push Forward
Google introduced Ironwood, a TPU designed specifically for the inference era. The chip was created using AlphaChip, an AI-driven design method.
The robotics team released Gemini Robotics and its successor Gemini Robotics 1.5, bringing AI agents into physical environments. Genie 3 launched as a new frontier for general-purpose world models.
Global Applications Expand
Weather forecasting saw significant advances. WeatherNext 2 now generates forecasts 8x faster with resolution up to 1-hour intervals. Flood forecasting information covers more than two billion people across 150 countries.
FireSat launched in March to detect wildfires earlier. AlphaEarth Foundations began mapping the planet in unprecedented detail.
Safety Frameworks Strengthen
Google positioned Gemini 3 as its most secure model yet, having undergone the most comprehensive safety evaluations of any Google AI model. The company published its approach to responsible AGI development in April, emphasizing readiness, proactive risk assessment, and industry collaboration.
New verification tools arrived for AI-generated images and videos within the Gemini app.
Industry Collaboration Grows
Google helped form the Agentic AI Foundation alongside other major AI labs to support open standards for agentic AI development. The company announced Model Context Protocol support for Google services in December.
Research partnerships with universities including UC Berkeley, Yale, and University of Chicago contributed to frontier research. Google also began working with the U.S. Department of Energy’s 17 national laboratories through the Genesis initiative.
