Gemini 3 Pro: Google's New AI Model Aims to Redefine Multimodal Understanding

What you need to know

• Gemini 3 Pro reclaims AI benchmarks, outperforming previous models and dominating both LMArena and WebDev Arena leaderboards.
• New AI mode enhances Google Search with multimodal responses, offering smarter, concise interactions for users.
• Gemini 3 introduces breakthrough features in Google Antigravity, enabling advanced agentic development and independent coding capabilities.

In a landmark release, Google has unveiled Gemini 3, its most advanced AI model yet, now available across various platforms. With top scores on major benchmarks, Gemini 3 Pro promises unprecedented reasoning and multimodal capabilities, redefining user interaction.

The Gemini 3 rollout marks the first time Google is shipping a new AI model in Search on day one. AI Mode in Search users can leverage the model's better reasoning and agentic coding capabilities to get help with tougher questions, while creating multimodal responses that include text and interactive visuals. It's also arriving in the Gemini app, bringing a Gemini Agent that can handle multi-step tasks independently.

Gemini 3 launches with Gemini 3 Pro in preview, plus a sneak peek at Gemini 3 Deep Think, an even smarter AI model that will come to Google AI Ultra subscribers in the coming weeks. Google says more Gemini 3 versions will be released "soon" as the company builds out its next-generation suite of AI models.

Gemini 3 reclaims the top spot on the LMArena and WebDev Arena leaderboards

Google's Gemini 2.5 models, released in March 2025, performed well on top AI benchmarks for months before losing ground to xAI and Anthropic models of late. Gemini 3 Pro flips that trend, as the AI model leads both the LMArena and WebDev Arena leaderboards immediately following its launch. Compared to Gemini 2.5 Pro, the new model significantly outperforms it on every major AI benchmark.

Notably, Gemini 3 Pro tops the LMArena leaderboard with a score of 1,501 points, dethroning Grok 4.1 Thinking, which earned a score of 1,483 points. For perspective, the Gemini 2.5 Pro model slotted in at third place with 1,452 points before Gemini 3 Pro's release. In other benchmarks, like Humanity's Last Exam, Gemini 3 Pro scored 37.5% without tool use — a result that demonstrates PhD-level reasoning, according to Google.

For the average user, Gemini 3 crucially "brings a new level of depth and nuance to every interaction." AI models tend to have a people-pleasing nature to them. They tell users what they want to hear, not what is actually true. Gemini 3 Pro responses are "smart, concise, and direct" by comparison, aiming to dispel the notion that AI models can't tell users what they need to hear.

For website creation and generative user-interface design, Gemini 3 is "the best vibe coding and agentic coding model" Google says it has ever built. That's demonstrated by a leading 1,487 ELO score on the WebDev Arena leaderboard, and better performance on Terminal-Bench 2.0 and SWE-bench Verified benchmarks. The result is a Gemini 3 Pro model that can build playable 3D video games from scratch, create websites, or fulfill your vibe-coding needs.

For better performance, there's Gemini 3 Deep Think, improving results on Humanity's Last Exam to 41% without tool use. It pushes Gemini 3's reasoning and multimodal understanding capabilities further, helping users solve the most complex of problems. This model is currently undergoing extra safety evaluations and will reach safety testers before it hits Google AI Ultra subscribers in the coming weeks.

Gemini 3 is now live in Google AI Studio, Vertex AI, Gemini CLI, and Antigravity

Gemini 3 is not only widely rolling out in Google's consumer products, it's also becoming available for developers. It's available today in Google AI Studio, Vertex AI, and Gemini CLI, as well as third-party sites like Cursor, GitHub, JetBrains, Manus, and Replit. Most importantly, Gemini 3 is the foundation for Google Antigravity, a brand-new agentic development platform.

Google Antigravity is essentially an AI-powered integrated development environment (IDE). This isn't new, but Antigravity's smarter AI agents separates itself from every other IDE. Gemini 3's new features, like advanced reasoning and better agentic coding capabilities, form AI agents in Antigravity that have unprecedented permissions.

AI agents have direct access to the editor, terminal, and browser in Google Antigravity, allowing them to become coding partners that can plan, execute, and validate software tasks independently. The platform uses a mix of Gemini 3 Pro, Gemini 2.5 Computer Use, and Nano Banana (Gemini 2.5 Image) to enable end-to-end agentic workflows.

Try Gemini 3 now in the Gemini app and Google Search

Gemini 3 Pro is rolling out to Google's consumer products, including Search and the Gemini app. Google Search users subscribed to Google AI Pro or Ultra can try it by selecting the "Thinking" model in AI Mode. It'll soon come to all AI Mode in Search users in the U.S., but for now, it's a paid exclusive.

In the coming weeks, AI Pro and Ultra users will get a new automatic model selector in Search that routes challenging requests to flagship Gemini 3 models, while sending simpler tasks to quicker and more efficient models.

Gemini 3's better understanding allows AI Mode to use the query fan-out method to perform more searches at once and uncover new content based on your intent. By correctly understanding the purpose of a user's request, Gemini 3 and AI Mode can surface relevant information that it previously would've ignored on older models.

AI Mode also uses generative UI to create interactive and engaging learning tools in responses to searches, like images, tables, and grids. It can go one step further to code a custom simulation or create a tool in real-time, if your prompt requires it.

Gemini 3 Pro is available now in Google Search for subscribers and the Gemini app for all users.