19th November 2025 – (San Francisco) Google has launched Gemini 3, billed as its most capable model to date, unifying the suite’s multimodal strengths to turn ideas into working outputs with markedly improved reasoning, context understanding and intent recognition. The company is deploying the model at global Google scale from day one across Search’s AI Mode, Workspace Enterprise, Vertex AI and a new agentic development environment called Google Antigravity.
Chief executive Sundar Pichai said the Gemini era, begun nearly two years ago, has accelerated the delivery of advanced capabilities to billions via Google’s full‑stack approach, with AI Overviews now serving two billion monthly users, over 70% of Cloud customers adopting AI features and 13 million developers building on its generative platforms.
The first release in the new line, Gemini 3 Pro, enters preview with substantial benchmark gains over 2.5 Pro in both text and multimodal tasks. It leads the LMArena leaderboard with 1501 Elo, reaches 37.5% on Humanity’s Last Exam without tools and 91.9% on GPQA Diamond, and sets a new mathematical high-water mark of 23.4% on MathArena Apex. Its multimodal performance includes 81% on MMMU‑Pro, 87.6% on Video‑MMMU and 72.1% on SimpleQA Verified, signalling stronger factual reliability and breadth across science and maths. Responses are designed to be succinct and insight‑driven, translating dense material, generating code for high‑fidelity visualisations and assisting with creative ideation.
For developers, Google is introducing Antigravity, an agentic platform that elevates AI from assistant to active collaborator. Powered by Gemini 3’s reasoning, tool use and agentic coding, Antigravity provides agents with direct access to editor, terminal and browser, enabling autonomous planning, execution and validation of complex end‑to‑end software tasks. The platform integrates the Gemini 2.5 Computer Use model for browser control and the Nano Banana image editor (Gemini 2.5 Image), and is available in public preview at no charge on macOS, Linux and Windows.
Google is also piloting Gemini 3 Deep Think, an enhanced reasoning mode. In internal testing it surpasses Gemini 3 Pro with 41.0% on Humanity’s Last Exam (tool‑free), 93.8% on GPQA Diamond and 45.1% on ARC‑AGI‑2 with code execution (ARC Prize Verified), indicating stronger performance on novel problems. Access is being provided first to safety testers before a forthcoming release to Google AI Ultra subscribers.
The model’s practical footprint spans building, planning and learning. In software creation it leads WebDev Arena at 1487 Elo, posts 54.2% on Terminal‑Bench 2.0 for command‑line tool use and achieves 76.2% on SWE‑bench Verified for agentic coding. Developers can start in Vertex AI, Gemini CLI, Antigravity and through partner platforms including Cursor, GitHub, JetBrains, Manus and Replit. In long‑horizon planning it tops Vending‑Bench 2, maintaining consistent decisions and tool use over a simulated year of operating a vending business, an indicator of its ability to manage multi‑step workflows such as email triage or travel planning under user guidance. For learning, Gemini 3 combines state‑of‑the‑art reasoning, vision and spatial understanding with multilingual strength and a one‑million‑token context window to synthesise text, imagery, video, audio and code, transforming dense research into interactive flashcards, visualisations and targeted training plans.
Search receives a notable upgrade through Gemini 3. Google’s query fan‑out is enhanced to run broader, intent‑aware explorations of the web, retrieving more relevant, high‑quality material. AI Mode now supports generative experiences such as immersive visual layouts, interactive tools and simulations, produced dynamically from the user’s query. From today, Google AI Pro and Ultra subscribers in the United States can select Gemini 3 Pro via the “Thinking” option in AI Mode, with automatic routing of tougher questions to Gemini 3 rolling out in the coming weeks while faster models handle simpler tasks.
Safety evaluations for Gemini 3 have been the most extensive yet at Google, with reduced sycophancy, stronger resistance to prompt injection and tightened defences against cyber‑enabled misuse. In addition to in‑house assessments under the Frontier Safety Framework, Google has collaborated with leading external experts, given early access to bodies such as the UK AISI and secured independent reviews from organisations including Apollo, Vaultis and Dreadnode.
The rollout begins immediately for Workspace Enterprise users of Gemini, developers via the Gemini API, Antigravity and the Gemini CLI, and enterprises through Vertex AI and Gemini Enterprise. Deep Think will follow after additional safety testing before becoming available to Google AI Ultra subscribers.
