Gemini 3 Arrives: Google Unveils ‘Antigravity’ Platform for Agentic AI Development
Google has announced the launch of Gemini 3, the next generation of its flagship artificial intelligence model, which the company is immediately integrating into Search, the Gemini app, cloud services, and developer tools. Google’s leadership calls Gemini 3 the most intelligent model in its portfolio and another step toward the long-sought goal of universal artificial intelligence, or AGI.
Nearly two years ago, the company inaugurated what it termed the Gemini era, and since then the scale of its AI usage has grown dramatically. According to Google, AI Overviews in Search now reach roughly 2 billion users each month, the Gemini app has surpassed 650 million monthly active users, more than 70 percent of Google Cloud customers already rely on its AI services, and around 13 million developers have experimented with its generative models.
Each generation of the Gemini family has built on the last. The first version introduced multimodal capabilities and support for long-context inputs. The second laid the foundation for agentic behavior and advanced reasoning. Gemini 2.5 then held the top position on the widely watched LMArena leaderboard for several months. Gemini 3 now unifies these achievements within a single core, aiming to better understand complex user queries, interpret context and intent, and behave more like a thoughtful and perceptive digital interlocutor.
Google asserts that Gemini 3 Pro delivers record-setting performance across numerous industry benchmarks in logic, mathematics, and factual accuracy, significantly surpassing the previous generation, 2.5 Pro. In evaluations such as GPQA Diamond and Humanity’s Last Exam, the model demonstrates reasoning at an expert level, and it sets new standards in specialized mathematical test suites. The company also highlights major advances in multimodal assessments involving simultaneous interpretation of text, images, and video.
Yet the developers emphasize not only “numbers on a chart,” but also how the model behaves in ordinary conversation. According to Google, Gemini 3 strives to answer concisely and purposefully, avoids empty pleasantries and formulaic phrasing, and instead focuses on providing candid, genuinely helpful responses that illuminate the topic or offer a fresh perspective.
One of the model’s defining qualities lies in its training paradigm. Gemini 3 was conceived from the outset as a fully multimodal system: it can process text, images, video, audio, and code at once, and its one-million-token context window allows it to work with exceptionally long materials. Google cites examples in which Gemini 3 “deciphers” old handwritten recipes, translates them from multiple languages, and compiles them into a family cookbook; transforms scientific papers and hours-long lectures into interactive study guides and flashcards; and analyzes sports-training footage, pointing out common errors and proposing structured improvement plans.
A second major application area involves software development. Google positions Gemini 3 as its strongest model to date for “vibe coding,” where a developer describes the desired outcome, and the AI handles much of the routine programming and interface assembly. The company claims that Gemini 3 leads benchmarks such as WebDev Arena and is markedly better than its predecessors at tasks requiring not just code generation, but correct use of tools, terminals, and external APIs.
To accompany Gemini 3’s launch, Google has introduced a new platform: Google Antigravity. This agent-oriented development environment elevates the AI from a “chat sidebar” to a first-class participant with direct access to the code editor, terminal, and integrated browser. The agent can plan its own workflow, break tasks into steps, run multiple processes in parallel, test and validate its code, and produce detailed artifacts—including plans, logs, and screenshots—so that the user can follow exactly what it has done. Antigravity employs not only Gemini 3 Pro, but also the specialized Gemini 2.5 Computer Use model for browser control, as well as Google’s Nano Banana engine for generating and editing images.
Gemini 3 is also designed to be a more dependable task executor in everyday life. Google reports that the model plans long-horizon actions more effectively, as evidenced by tests like Vending Bench, where the AI manages a virtual vending-machine business and must make economically sound decisions over the course of a simulated year. In practical scenarios, this translates into an ability to carry multi-step processes through to completion—such as booking services that require several checkpoints or untangling and organizing an overwhelmed email inbox—while relying on built-in tools and external services.
For the most intricate problems, Google is preparing a dedicated mode called Gemini 3 Deep Think. The company says it handles unconventional reasoning tasks more adeptly, especially those with no obvious correct answer. Deep Think is undergoing additional safety evaluations and is currently available only to a limited pool of testers; eventually, it will be offered to subscribers at the Google AI Ultra tier.
A substantial portion of the announcement centers on safety. Google states that Gemini 3 has undergone the most extensive evaluation cycle of any model in the company’s history, is more resistant to manipulative or harmful prompts, is less inclined to indulge user attempts at misuse, and is better protected against scenarios such as automated cyberattacks. To assess risks, Google enlisted not only internal teams but also external experts, including relevant UK government bodies and independent security firms that audited the system.
The rollout of Gemini 3 is already underway. The model is appearing in the Gemini app, in AI Mode within Search for Google AI Pro and Ultra subscribers, in developer tools such as the Gemini API in AI Studio and the Gemini CLI, and for enterprise customers through Vertex AI and the Gemini Enterprise suite. Certain agentic features—like the Gemini Agent for email management—are already available to advanced subscribers, while Deep Think capabilities will be added incrementally as safety reviews conclude.
Google now possesses not only new benchmark records but also unprecedented scale: the company is deploying the model directly into Search and its core products, which reach billions of users. The open question is how valuable this upgrade will prove for everyday users and developers—and how quickly Gemini 3 can evolve from a striking demonstration of AI sophistication into a truly indispensable working instrument.
Support Our Threat Intelligence
If you find our technology report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
