On November 18, 2025, Google unveiled Gemini 3, its latest and most advanced AI model to date. Developed by Google DeepMind, Gemini 3 marks a significant advancement in artificial intelligence, integrating multimodal understanding, agentic behavior, and cutting-edge reasoning within a single, cohesive system. Building upon the foundations set by its predecessors, Gemini 3 is now accessible across the Google ecosystem, including Search, the Gemini app, AI Studio, Vertex AI, and Google Antigravity.
Gemini 3 surpasses earlier models with its exceptional performance in reasoning and problem-solving. It achieves a 1501 Elo score on the LMArena Leaderboard and exhibits PhD-level reasoning, excelling in AI benchmarks such as Humanity’s Last Exam with a score of 37.5% and GPQA Diamond at 91.9%. Its multimodal intelligence shines with an 87.6% score on Video-MMMU and 81% on MMMU-Pro. Additionally, it achieves advanced factual accuracy with a 72.1% score on SimpleQA Verified, making it a leader in addressing complex queries across various formats, including text, image, video, audio, and code.
Google is also introducing Gemini 3 Deep Think, an enhanced mode that pushes the boundaries of AI reasoning. This mode records a 41.0% score on Humanity’s Last Exam, outperforming the Pro version, and achieves 93.8% on GPQA Diamond and 45.1% on ARC-AGI-2. These scores highlight its unprecedented ability to tackle novel problems through code execution. Following further safety evaluations, this mode will be available to Google AI Ultra subscribers.
Gemini 3 offers diverse applications, enabling users to learn from handwritten notes, academic texts, and complex videos. It transforms these inputs into interactive flashcards, shareable cookbooks, code-based visualizations, and personalized training plans. With support for 1 million-token context windows, it can perform deep analyses of extensive content seamlessly.
For developers, Gemini 3 is available on Google AI Studio, Google CLI, and other third-party platforms like GitHub, Replit, and JetBrains. It enhances developer productivity, leading the WebDev Arena leaderboard with a 1487 Elo score and performing well in command-line operations and agentic coding.
In planning, Gemini 3 excels in long-horizon tasks, outperforming peers on Vending-Bench 2 by simulating a year-long vending machine business with superior ROI and tool consistency. It assists with real-life planning, including booking appointments, managing schedules, and conducting multi-step tasks autonomously.
Google introduces Antigravity, a new development environment designed for agentic coding with Gemini 3. This platform allows agents to autonomously plan and execute comprehensive software development tasks, validate code, and run browser-based simulations, enabling end-to-end workflows with minimal human intervention.
Gemini 3 stands as Google’s most tested and secure AI model, equipped with enhanced safeguards against prompt injection attacks, biases, and misuse through cyber threats. Google collaborated with global safety experts and ran the model through its Frontier Safety Framework to ensure its robustness before rollout.
Starting November 18, 2025, Gemini 3 is being rolled out across Google Search (AI Mode), the Gemini app for Google AI Pro and Ultra subscribers, Google AI Studio, Vertex AI for enterprises, and Google Antigravity for developers. Gemini 3 Deep Think mode will be available following further safety validations.
Kutos : AI Assistant!
