Google launches Gemini 3 in Search, Gemini and Antigravity

Google has unveiled Gemini 3, the next generation of its artificial intelligence model family. The launch goes beyond a technical demo: Gemini 3 Pro is available today in preview through AI Mode in Search, the Gemini app, AI Studio and Vertex AI, Google Cloud’s platform.

The decision matters because of its reach. Google does not typically need to win over a new audience to distribute a model: it already has products used by hundreds of millions of people and businesses. Sundar Pichai has said that AI Overviews reaches 2 billion users each month and that the Gemini app has more than 650 million monthly users.

A model built to reason across text, images and video

Gemini 3 Pro is a multimodal model: it can work with text, images, audio, video and code within the same task. Google says it outperforms Gemini 2.5 Pro on the key benchmarks it uses to evaluate reasoning, coding and understanding across different formats.

On LMArena, a leaderboard based on user preferences when comparing model responses, Gemini 3 Pro scored 1,501 Elo points. The company also reports a 37.5% score on Humanity’s Last Exam without tools, a deliberately difficult test with questions spanning multiple disciplines, and 91.9% on GPQA Diamond, which focuses on advanced scientific questions.

Multimodal results are another strength of the launch. Gemini 3 Pro scores 81% on MMMU-Pro, which combines visual and knowledge-based problems, and 87.6% on Video-MMMU. On SimpleQA Verified, a factual-accuracy evaluation, it reaches 72.1%.

These figures are useful for measuring progress between models, but they do not guarantee correct answers for every query. A model can excel on standardized exams and still require human review when used for professional decisions, changing data or high-impact information.

Search gets Gemini 3 from day one

The most relevant change for the general public is its immediate integration into AI Mode, Google’s conversational search mode. The company says the model will handle more complex queries and create dynamic experiences within Search.

That is a significant distribution shift. Earlier generations of frontier models typically reached users first through an API, a standalone app or a waitlist. Google is bringing Gemini 3 to its flagship product at the same time it makes it available to developers.

For users, the promise is to reduce the number of instructions they need to give. Instead of separately asking for a document summary, an explanation of a chart and a study guide, the model aims to understand the entire set of materials and produce a tailored response. Google cites translating handwritten recipes, analyzing academic papers and creating visualizations and study guides from long videos as examples.

Gemini 3 also retains a one-million-token context window. Put simply, it can take a very large amount of text, code or audiovisual material into account during a conversation. That does not mean it will always interpret a million tokens with the same level of reliability, but it does expand the kinds of documents it can process without breaking them into small chunks.

Deep Think and a new platform for agentic coding

Google has also announced Gemini 3 Deep Think, an enhanced reasoning mode for especially complex problems. Before making it available to Google AI Ultra subscribers, the company will provide it to safety testers.

In its internal tests, Deep Think scores 41% on Humanity’s Last Exam without tools, 93.8% on GPQA Diamond and 45.1% on ARC-AGI-2 with code execution. ARC-AGI-2 measures the ability to infer new rules from a small number of examples, an area where current models tend to struggle more than they do with familiar questions.

The launch also includes Google Antigravity, an agentic development platform. A coding agent does more than suggest lines of code: it can plan tasks, use tools and execute steps to complete a development goal. Google presents Gemini 3 as its most capable model yet for these workflows and for so-called vibe coding, in which users describe the result they want in natural language.

For developers and businesses, Gemini 3’s availability today in AI Studio and Vertex AI makes it easier to test the model in both prototypes and cloud environments. The question will be whether its lab results translate to real-world tasks: software maintenance, analysis of extensive documentation, process automation and products that combine voice, image and text.

Google is thus opening a new chapter for Gemini with a clear strategy: rather than separating research from the product, it is putting the model in Search, its apps and its professional tools from day one.