Google has unveiled Gemini 3, its latest flagship AI model, boasting significant improvements across the board over its predecessor, Gemini 2.5. The company says early benchmark results from LM Arena support its claims.
The new model is rolling out to consumers, developers, and enterprise customers, alongside a new platform for building AI agents called Antigravity.
A More Capable, Context-Aware Model
Google is positioning Gemini 3 as a general-purpose system with sharper reasoning skills and deeper comprehension of natural-language programming, often referred to as “vibe coding.” Executives, including CEO Sundar Pichai, say the model handles context and intent more effectively and now interprets spatial information in addition to text and images.
While Google has not disclosed architectural specifics or details about the training data, it states that improvements span multiple components of the system.
Redesigned App With Magazine-Style Responses
Gemini 3 is now integrated into the updated Gemini app, featuring a redesigned interface that presents answers in a polished, magazine-like format. In demonstrations, Google showed the app generating travel itineraries and narrative-style content about Vincent van Gogh.
The company cautions that despite the slick presentation, responses may still contain inaccuracies. A specialised variant, Gemini 3 Deep Think, optimised for advanced reasoning, is in limited testing with security researchers and is currently available only to Google AI Ultra subscribers.
Benchmark Results
According to Google, Gemini 3 Pro now leads the LM Arena leaderboard with a score of 1501. On Humane’s Last Exam benchmark, Gemini 3 Pro reached 37.5%, up from 21.6% for Gemini 2.5 Pro. The Deep Think version scored about 41% on the same test.
AI Mode Coming to Search
Google is also preparing to bring Gemini 3 into its search ecosystem through a new feature called AI Mode. For U.S. subscribers to Google AI Pro and Ultra, an advanced “Thinking” option powered by Gemini 3 Pro will soon be available.
AI Mode will introduce a generative interface capable of moving long text directly into the prompt bar, producing interactive layouts, and running simulations — features made possible by the model’s agent-like capabilities.
Antigravity: Google’s New Agent-Building Platform
Alongside the model launch, Google introduced Antigravity, a developer-focused environment for building and managing AI agents. Functioning more like an integrated development environment (IDE) than a consumer assistant, Antigravity provides access to an editor, terminal, and browser.
The platform features a browser-driven agent called Gemini 2.5 Computer Use, which can interact with websites using screenshots, as well as tools such as Nano Banana for image generation.
Developers can use Gemini 3 through the API in AI Studio, Antigravity, and the Gemini CLI. Enterprise customers will gain access via Vertex AI and Gemini Enterprise.
Google says Gemini 3 is its safest model yet, with strengthened protections against prompt-injection attacks and misuse scenarios, including those linked to cyberattacks.
