Gemini 3 goes live with big upgrades in reasoning, coding and search features

New Delhi: Google’s next major AI chapter is officially here. The company has launched Gemini 3, its most advanced model yet, only a day after we broke the news about Gemini 3.0 Pro’s benchmark scores and model card appearing online ahead of schedule. The early leak had already hinted at a significant jump in reasoning and multimodal capability. Today’s announcement confirms that Google is pushing hard to keep pace in the global AI race.

The company is rolling Gemini 3 directly into several of its most widely used products, including Search, the Gemini app and enterprise tools on Vertex AI. This is one of the fastest deployment cycles Google has ever attempted for a flagship model. I felt a similar sense of speed only once earlier this year when the Gemini app crossed hundreds of millions of users in a matter of weeks. This update feels like Google trying to match that pace on the model side.

Google and Alphabet CEO Sundar Pichai set the tone in the blogpost, calling Gemini 3 "our most intelligent model.” He added that the company can now get advanced capabilities to users faster because of its "differentiated full stack approach to AI innovation.” Pichai also said the launch marks the first time Google has shipped a new model into Search on day one.

The company credits Gemini 1 and Gemini 2 for building the base for this moment. Pichai pointed out that Gemini 2.5 Pro topped LMArena for over six months and set the path for deeper reasoning features. Now Gemini 3 aims to bring all those abilities together so users can "bring any idea to life.”

Demis Hassabis, CEO of Google DeepMind, and Koray Kavukcuoglu, DeepMind’s CTO, said the release is "another big step on the path toward AGI.” They announced that Gemini 3 Pro will be available in preview from today across Google products.

Gemini 3 Pro improves on nearly every benchmark. Google says it scored 1501 Elo on the LMArena leaderboard and hit 91.9 percent on GPQA Diamond. It also posted 37.5 percent on Humanity’s Last Exam without tools and topped several multimodal tests, including 81 percent on MMMU Pro and 87.6 percent on Video MMMU.

These are dense numbers, but the simple takeaway is that Gemini 3 handles long, complex queries better. The model’s responses feel more direct and less fluffy. In fact, Google says it avoids clichés and focuses on useful insight. This part caught my attention because I have seen how easily AI models slip into overly polite or repetitive tones when confused. Gemini 3 is supposed to hold its ground better.

Google also introduced Gemini 3 Deep Think. Early testing shows even higher scores, including 45.1 percent on ARC AGI 2 with code execution. The company will give access to safety testers first before releasing it to Google AI Ultra subscribers.

The biggest everyday impact will be in Google Search. AI Mode now uses Gemini 3 to offer visual layouts, interactive tools and dynamic simulations. If someone searches for RNA polymerase, for example, Search can create a generated interface with visuals and step by step explanations.

The Gemini app is getting a redesign as well. Google says the app can now respond to prompts with UI that feels like a small website. One example from the briefing showed Gemini generating a Van Gogh gallery with interactive elements. I have personally seen many friends use the app for daily planning, so a richer interface might make it feel more like a full assistant.

Google also launched Gemini Agent. It can complete multi step tasks such as organising email inboxes or booking travel. Pichai described this as the first step toward a broader "universal assistant” vision.

The enterprise announcement came in a separate blogpost. Google said Gemini 3 is now in Gemini Enterprise and Vertex AI. The company highlighted three themes for businesses • state of the art reasoning and multimodal understanding • powerful agentic coding • advanced tool use and planning

Ben Kus, CTO of Box, said Gemini 3 Pro can transform workflows across sales, marketing, legal and finance by improving how companies analyse their own content. Presentations.ai CEO Sumanth Raghavendra said Gemini 3 can turn six hours of analysis into 90 seconds of output.

Google also showed examples from Rakuten, GitHub, JetBrains, Wayfair, Figma and Shopify, all praising the model’s improved reliability. Koray Kavukcuoglu said the company is releasing models faster and getting them to users earlier than ever.

Google says Gemini 3 is its "most secure model yet.” It has undergone extensive evaluations with partners like the UK AISI and external experts including Apollo, Vaultis and Dreadnode. Pichai said the next era of Gemini starts now, with the rollout beginning today across Search, the Gemini app, Enterprise, Vertex AI, AI Studio, Google Antigravity and Gemini CLI.

The company will add more models to the Gemini 3 series soon. For now, the flagship Gemini 3 Pro and the preview of Deep Think signal one of the biggest steps Google has taken since launching the first Gemini model two years ago.