Google gets closer to AGI with Gemini 3
Gemini 3 represents a major step toward artificial general intelligence (AGI) by combining state-of-the-art reasoning, multimodal understanding, and agentic abilities, enabling users to "bring any idea to life" through learning, building, and planning tasks at scale.
Google unveiled Gemini 3 on November 18, 2025, describing it as its most advanced AI model yet, integrating enhanced reasoning, multimodal processing (text, images, video, audio, code), and agentic capabilities for autonomous task execution. The model includes variants like Gemini 3 Pro, which outperforms its predecessor Gemini 2.5 Pro on benchmarks such as LMArena (1501 Elo), Humanity’s Last Exam (37.5%), GPQA Diamond (91.9%), MMMU-Pro (81%), and SWE-bench Verified (76.2%). A new "Deep Think" mode further boosts performance on complex tasks, achieving scores like 41.0% on Humanity’s Last Exam and 45.1% on ARC-AGI-2. Key features encompass a 1 million-token context window, multilingual support, zero-shot generation, and tools like Google Antigravity for agentic workflows. Availability begins immediately for Gemini 3 Pro in the Gemini app, AI Studio, Vertex AI, and third-party platforms, with enterprise access via Gemini Enterprise; Deep Think mode is in safety testing for Ultra subscribers. The announcement highlights integrations with Google products, such as AI Mode in Search for generative interfaces and agentic features in the Gemini app for tasks like inbox management. Safety measures include evaluations under the Frontier Safety Framework to mitigate risks like sycophancy and prompt injections.
Why Gemini 3 Is a Big Deal
Gemini 3 stands out due to its benchmark-leading performance across reasoning, multimodal, and coding tasks, surpassing previous models and competitors in areas like PhD-level reasoning (GPQA Diamond) and long-horizon planning (Vending-Bench 2). It advances agentic AI with tools like Antigravity for autonomous operations, potentially transforming workflows in development, education, and enterprise settings. The immediate global rollout across Google ecosystems and third-party integrations signals rapid deployment potential, while its focus on safety and reliability addresses ongoing AI concerns. As noted by Google executives, it builds on two years of Gemini development, positioning it as a foundational release for broader AI adoption and innovation.
Google's Approach to Achieving Gemini 3 Performance Levels
Google's Gemini 3 model, released on November 20, 2025, demonstrates enhanced performance across reasoning, multimodal processing, and agentic tasks, as outlined in its official announcement and subsequent analyses. The model builds on prior iterations, incorporating architectural refinements and training optimizations to outperform predecessors like Gemini 2.5 Pro on benchmarks such as LMArena (1501 Elo), GPQA Diamond (91.9%), and SWE-bench Verified (76.2%).
Key methods contributing to these gains include:
- Sparse Mixture of Experts (MoE) Architecture: This design enables efficient scaling by activating only relevant expert sub-networks during inference, reducing computational costs while maintaining high performance; it supports a 1 million-token input context and 64k-token output, allowing for complex, long-horizon tasks at approximately 40% lower cost compared to competitors like Claude Sonnet 4.5.
- Advanced Pre- and Post-Training Integration: Gemini 3 leverages scaling laws through seamless pre-training (large-scale data ingestion) and post-training phases, including reinforcement learning (RL) specifically for tool-use, resulting in improved factual accuracy, reasoning, and agentic capabilities for autonomous task execution.
- JAX Framework and TPU Hardware: Training utilized JAX, a high-performance machine learning library with just-in-time compilation and XLA optimization, enabling faster execution across hardware; the model was trained exclusively on Google's seventh-generation TPUs (Ironwood), offering a 10x peak performance boost over previous versions for both training and inference.
- Deep Think Mode: An innovative reasoning enhancement that generates parallel hypothesis-testing structures, elevating scores on challenging benchmarks like Humanity’s Last Exam (41.0%) and ARC-AGI-2 (45.1% with code execution), allowing the model to self-refine answers dynamically.
- Multimodal and Agentic Innovations: Native multimodality from earlier Gemini models is deepened for better contextual understanding in video and images (e.g., 87.6% on Video-MMMU); agentic features, powered by tools like Google Antigravity, enable end-to-end task planning and execution, such as software development or cybersecurity challenges.
These elements collectively position Gemini 3 as a frontier model, with safety evaluations under Google's Frontier Safety Framework ensuring reliability against issues like prompt injections. While specifics on training data remain undisclosed, the focus on infrastructure moats, like proprietary TPUs and distribution channels, underscores Google's strategic emphasis on scalable AI development.
Key Figures Behind Gemini 3's Development
Google's Gemini 3, announced on November 18, 2025, represents a collaborative effort led by top executives and AI specialists within Google and Google DeepMind. The following individuals stand out based on their roles in the model's announcement, oversight, and technical direction:
- Sundar Pichai: As CEO of Google and Alphabet, Pichai provided strategic oversight and publicly introduced Gemini 3, emphasizing its advancements in multimodal AI and agentic capabilities.
- Demis Hassabis: CEO of Google DeepMind, Hassabis co-authored the launch announcement and has been instrumental in guiding the model's research, drawing from DeepMind's expertise in advanced AI systems.
- Koray Kavukcuoglu: Serving as CTO of Google DeepMind and Chief AI Architect at Google, Kavukcuoglu co-authored the Gemini 3 announcement on behalf of the development team, contributing to its architectural innovations.
- Josh Woodward: A senior product leader at Google DeepMind, Woodward discussed Gemini 3's features in public segments, highlighting its practical applications.
The project also benefited from broader team efforts at Google DeepMind, including recent hires like Varun Mohan and Douglas Chen from AI startup Windsurf, who joined to enhance agentic coding aspects, though their direct impact on Gemini 3 remains unspecified. No additional individual contributors were named in the official release, which credited the "Gemini team" collectively.
References
-
Official Google Blog Announcement: This primary source details the model's release and is co-authored by Demis Hassabis (CEO, Google DeepMind) and Koray Kavukcuoglu (CTO, Google DeepMind), highlighting their roles in the project's success.
-
Gemini 3 Developer Guide: Provides technical insights into the model's features, reflecting contributions from Google DeepMind's AI architects and engineers.
-
YouTube Walkthrough with Google DeepMind Representative: Features Logan Kilpatrick from Google DeepMind demonstrating Gemini 3, offering direct commentary from a key team member involved in its development.