Google launches more powerful Gemini 3 model for reasoning and vibe coding

Since launching Gemini two years ago, Google’s AI chatbot has entered the mainstream. The Gemini app has over 650 million users, while AI Overviews are used by 2 billion people every month.

What can Gemini 3 do?

Google has highlighted that Gemini 3 is designed for strong multimodal understanding and is a powerful model for “vibe coding.” The company claims that Gemini 3 outperforms its predecessor on every AI benchmark, topping the LM Arena leaderboard with a score of 1501 points, which demonstrates PhD-level reasoning, earning top marks on Humanity’s Last Exam and GPQA Diamond.

Gemini 3 Pro also excelled in multimodal reasoning, scoring 81 per cent on MMMU-Pro and 87.6 per cent on Video-MMMU, showcasing how Google’s frontier model can solve complex problems across various topics, including science and mathematics.

During a briefing, Google demonstrated how Gemini 3 understands both visual and spatial information, and then delivers results. For example, Gemini 3 can decipher and translate handwritten recipes in different languages into a shareable cookbook. By providing Gemini with academic papers and video lectures, it can generate code to create interactive flashcards, making it easier and more visually engaging to learn new topics.

For developers, Google markets Gemini 3 as the best model for vibe coding and as its most advanced agentic coding model, outperforming the previous Gemini 2.5 Pro. Developers can build with Gemini 3 in Google AI Studio, Vertex AI, Gemini CLI, and Google Antigravity, the company’s new agentic development platform.

Google is calling Gemini 3 its most secure AI model, noting that it has undergone extensive safety evaluations. “The model shows reduced sycophancy, increased resistance to prompt injection, and improved protection against misuse via cyberattacks,” the company said.

Gemini brings a new type of interface to the Gemini app. (Image credit: Google)

New interface, Dynamic View

Users will also notice that the Gemini app has been cleaned up, and redesigned with a more modern look. Google has introduced what it calls “generative interfaces,” a new type of interface that appears dynamically when prompted. For example, if you ask Gemini to plan a “three-day trip to Rome next summer,” you will receive a visual itinerary similar to a magazine layout, complete with photos and modules.

Google is also adding another layer to the experience called Dynamic View, which designs and codes a custom user interface in real time. For instance, if you ask Gemini to “explain the Van Gogh gallery with life context for each piece,” you will get an interactive response that you can tap and scroll through.

New to the Gemini app is Gemini Agent, powered by Gemini 3. It is a tool that handles multi-step tasks directly within the app – it can connect to your Google apps, manage your calendar, add reminders, prioritise to-dos, and even draft replies. Gemini Agent is currently rolling out to Gemini AI Ultra users first.

Google is also bringing Gemini 3 to Search, starting with AI Mode. By selecting “Thinking” from the model drop-down menu in AI Mode, Gemini 3 enables users to ask complex and nuanced questions, including lengthy ones, allowing the AI to better understand the context and intent before responding. Gemini 3 in AI Mode is currently rolling out to all users in the US, with higher limits available to those with Google AI Pro and Ultra plans.

Google also announced a new agent platform called “Google Antigravity,” which lets developers code “at a higher, task-oriented level.”

Aggressive AI race

The year has been ruthless when it comes to AI dominance, and every major AI company is openly challenging its rivals to build better and faster LLMs.

Google is joining the race, following in the footsteps of OpenAI and Anthropic, which launched their own cutting-edge models just a couple of months ago. It’s unclear whether this combination of speed, power, and features will be enough, however. And while Gemini 3 is touted as a powerful model with strong multimodal capabilities and reasoning skills, Google still faces pressure from both new and established players as the market fills with frontier models that are more capable at reasoning and increasingly suited for autonomous tool use.

The bigger question is who will ultimately dominate the AI market, and whether that leadership will come from the US, China, or somewhere else. In January, Chinese upstart DeepSeek shocked the world with its open source model, showing a growing wave of powerful Chinese models now vying for global influence. Later, Mark Zuckerberg’s Meta formed a new superintelligence team that has lured away several top OpenAI scientists, although nobody knows exactly what the company is trying to achieve.