Google Releases Gemini 3 and Agent Platform

Gemini 3 is better than its predecessor, Gemini 2.5, on all levels—according to Google, which also cites results from the LM Arena, where AI models compete against each other. The new version is now being made available to consumers, developers, and businesses. Additionally, Google will offer Antigravity, a platform for creating and managing AI agents.

With Gemini 3, everyone should be able to bring any idea to life. The model is said to be particularly good at reasoning, i.e., the thought process, as well as Vibe Coding, which is how programming is described when natural language instructions are executed by AI. Furthermore, Gemini 3 is expected to better understand the context and intent of a request. In a blog post, Google CEO Sundar Pichai writes that it is exciting how AI can now not only read text and images but also understand space. What Google does not explain and answers evasively even when asked is how exactly the model has been improved. Is there a new architecture, different training data, or more data used? Improvements have been made everywhere, it is said in a conversation.

Gemini App: Answers in Magazine Style

Now Gemini 3 is integrated into the Gemini App. The responsible Google VP, Josh Woodward, says everyone will immediately notice how much more helpful Gemini has become with the update. In a press demonstration, he shows how Gemini can reorder the output. The entire look has been revamped. As an example, as with almost every AI presentation, the topic of travel planning is used. The app provides information almost in the style of a glossy magazine. Woodward also shows a storytelling about Vincent Van Gogh—also colorful, dynamic, and chic. Of course, the question always remains whether all information is correct, whether there are hallucinations, and where the information comes from. AI is still an experiment at Google.

Thus, Gemini 3 Deep Think, a model specialized in reasoning, is initially still being tested by security researchers and is only available to Google AI Ultra subscribers.

According to Google's Deepmind CEO, Demis Hassabis, Gemini 3 Pro outperforms all other models in the LM Arena, with 1501 points in the leaderboard. In Humane's Last Exam, for example, it achieves 37.5 percent (Gemini 2.5 Pro answered 21.6 percent correctly). Gemini 3 Deep Think even reaches 41 percent.

AI Mode Relies on Gemini 3

The search will also rely on Gemini 3 in the future, referring to the AI Mode. For Google AI Pro and Ultra subscribers, initially in the USA, there is an extension called Thinking, which uses Gemini 3 Pro. Google states that the AI mode will receive a new, generative user interface. This means that entire passages can be moved to the prompt bar, for example, as well as interactive layouts and simulations created by the AI model. Gemini also accesses its own agentic capabilities for this.

To create and manage AI agents, there is now a new platform: Google Antigravity. It is aimed at developers. So it's not about an agent that quickly buys a pair of socks for the user from the chat, as the ChatGPT Agent is supposed to do, for example. Fundamentally, it is an integrated development environment (IDE). The agents have access to the editor, terminal, and browser. Gemini 2.5 Computer Use integrated is also integrated into Antigravity. This is the AI agent that can operate the browser and uses screenshots for it. Nano Banana is also included as an image generator on the platform.

Developers have access to Gemini 3 via API in AI Studio, Antigravity, and Gemini CLI. For businesses, Gemini 3 is available via Vertex AI and Gemini Enterprise.

With all of this, Gemini 3 is said to be Google's most secure model to date. For example, it is less vulnerable to prompt injections or misuse for cyberattacks.

(emw)