Google introduced Gemini 3 on Tuesday, marking what the company calls a significant leap forward in AI reasoning, multimodal understanding, and agentic capabilities. The launch represents the next phase of Google’s two-year Gemini initiative, an effort that CEO Sundar Pichai described as one of the company’s most significant scientific and product undertakings. Pichai noted that the Gemini ecosystem has expanded rapidly, with AI Overviews reaching 2 billion monthly users and the Gemini app surpassing 650 million monthly users. Google reports that more than 70% of its Cloud customers now use its AI tools, and over 13 million developers have built with Gemini models.
Gemini 3, now available across Google products on launch day, is designed to unify and expand the capabilities introduced in earlier generations. While Gemini 1 focused on multimodality and expanded context, and Gemini 2 advanced agentic behavior and reasoning, Gemini 3 aims to deliver stronger contextual understanding, deeper planning abilities, enhanced multimodal interaction, and more autonomous coding and task execution. Google is positioning the model as a significant evolution, with Pichai emphasizing its ability to interpret intent with fewer prompts and provide more nuanced, insight-driven responses.
The company claims that Gemini 3 Pro, released in preview, outperforms its predecessor across all major benchmarks. It reached a record 1501 Elo on the LMArena Leaderboard and delivered high scores on advanced reasoning evaluations, such as Humanity’s Last Exam at 37.5% without tools and GPQA Diamond at 91.9%. It achieved a new state-of-the-art 23.4% on MathArena Apex. In multimodal evaluations, it posted 81% on MMMU-Pro and 87.6% on Video-MMMU, while surpassing previous frontier models with a 72.1% score on SimpleQA Verified for factual accuracy.
Google also unveiled Gemini 3 Deep Think, an enhanced reasoning mode built for the most complex tasks. In testing, Deep Think topped the base model’s performance with 41.0% on Humanity’s Last Exam, 93.8% on GPQA Diamond, and an unprecedented 45.1% on ARC-AGI-2 with code execution, verified by the ARC Prize team. The company describes Deep Think as a step-change in problem-solving, with greater depth and reliability across novel challenges.
Gemini 3 is now integrated into AI Mode in Google Search, enabling dynamic generative interfaces that produce visually immersive explanations, simulations and layouts based on complex queries. The model is also live in the Gemini app, AI Studio, Vertex AI, and the Gemini API. Google highlighted several consumer and enterprise use cases, emphasizing the model’s blend of reasoning, spatial understanding, and multimodal analysis. It can digitize and translate handwritten family recipes into a publication-ready cookbook, interpret academic material, generate code-based learning aids or visualizations, and analyze sports footage, such as pickleball matches, to produce targeted performance insights.
For developers, Gemini 3 expands zero-shot generation and interactive UI creation. It leads the WebDev Arena leaderboard with a 1487 Elo score and achieved 54.2% on Terminal-Bench 2.0, which evaluates a model’s ability to use computer tools. It also recorded 76.2% on SWE-bench Verified, significantly outperforming Gemini 2.5 Pro. Alongside the model release, Google announced Google Antigravity, a new agent-first development environment built around Gemini 3.
The platform elevates AI agents into active development partners that can autonomously plan and execute multi-step software tasks using integrated access to the editor, terminal, and browser. Antigravity includes the Gemini 2.5 Computer Use model for browser operations and Google’s Nano Banana image editing system.
Google also emphasized significant progress in long-horizon planning. Gemini 3 topped the Vending-Bench 2 benchmark, maintaining stable decision-making across a simulated year of operations. These improved planning and tool-use abilities enable more sophisticated real-world workflows such as booking local services, navigating multi-step tasks or organizing large volumes of email. Google AI Ultra subscribers can use these capabilities through Gemini Agent in the Gemini app.
The company said Gemini 3 is its most secure model yet, undergoing the most extensive safety evaluation of any Google AI system. The model is said to have reduced sycophancy, increased resistance to prompt injection attacks and stronger safeguards against cyber misuse. Google conducted external evaluations with organizations such as the UK AISI, Apollo, Vaultis, and Dreadnode, and applied its internal Frontier Safety Framework. The company released a public model card outlining additional protections and findings.
Beginning today, Gemini 3 is rolling out to consumers through the Gemini app, to Google AI Pro and Ultra subscribers in AI Mode in Search, to developers through the Gemini API, AI Studio, Google Antigravity, and the Gemini CLI, and to enterprises through Vertex AI and Gemini Enterprise. Gemini 3 Deep Think will undergo additional safety review before becoming available to Google AI Ultra subscribers in the coming weeks. Google plans to expand the Gemini 3 family with additional models soon and describes the release as the start of a new era aimed at delivering more capable reasoning, more autonomous agent,s and more personalized AI experiences.
