Google has officially launched Gemini 3. The AI model combines strong reasoning abilities with multimodal capabilities and sets the bar high for development.
Gemini 3 Pro outperforms its predecessor, the 2.5 Pro, across all key AI benchmarks. With a score of 1501 Elo on the LMArena Leaderboard, the model sets a new standard. On Humanity’s Last Exam, it scores 37.5 percent without using tools, demonstrating PhD-level reasoning. The GPQA Diamond score is 91.9 percent.
In the field of mathematics, Gemini 3 Pro also sets a new benchmark with 23.4 percent on MathArena Apex. For multimodal reasoning, the model scores 81 percent on MMMU-Pro and 87.6 percent on Video-MMMU. Factual accuracy reaches 72.1 percent on SimpleQA Verified.
According to Google, the model’s response has become more direct and concrete. It provides insights rather than flattery. Users get what they need to hear, not just what they want to hear. The model functions as a true thinking partner, translating complex scientific concepts into code for visualizations or supporting creative brainstorming sessions.
Two years of Gemini development
Google started the Gemini project almost two years ago. Its impact within the Google ecosystem is now significant. AI Overviews reaches 2 billion users monthly, while the Gemini app has over 650 million active users. More than 70 percent of Google Cloud customers now use AI functionality, and 13 million developers are building with the company’s generative models.
Each generation builds on the previous one. Gemini 1 introduced native multimodality and an expanded context window. Gemini 2 laid the foundation for agentic capabilities and set the standard for reasoning. Gemini 2.5 Pro dominated the LMArena rankings for six months. Now Gemini 3 combines all these capabilities into a single intelligent system.
Google DeepMind CEO Demis Hassabis and CTO Koray Kavukcuoglu call it “the world’s best multimodal model for understanding and our most powerful agentic and vibe coding model to date.”
Deep Think mode for complex problems
Gemini 3 Deep Think mode goes beyond the standard version. This enhanced reasoning mode performs better on Humanity’s Last Exam (41.0 percent without tools) and on GPQA Diamond (93.8 percent). On ARC-AGI-2, it achieves an unprecedented 45.1 percent with code execution.
Deep Think mode will undergo additional safety testing before becoming available to Google AI Ultra subscribers in the coming weeks.
Developers get new capabilities
For developers, Gemini 3 offers improved capabilities for zero-shot generation. The model processes complex prompts and instructions to render richer, more interactive web UIs. Google calls it its best-ever coding and agentic coding model.
On the WebDev Arena ranking, Gemini 3 achieves a score of 1487 Elo. Terminal-Bench 2.0 shows a score of 54.2 percent, which measures the tool’s ability to operate a computer via terminal. On SWE-bench Verified, the model scores 76.2 percent, significantly higher than 2.5 Pro.
Developers can now build with Gemini 3 in Google AI Studio, Vertex AI, Gemini CLI, and the new agentic development platform Google Antigravity. The model is also available through platforms such as Cursor, GitHub, JetBrains, Manus, and Replit.
Antigravity changes the developer experience
Google is simultaneously introducing Google Antigravity, a new agentic development platform. This platform enables developers to work at a higher, task-oriented level. The agents have direct access to the editor, terminal, and browser.
They can autonomously plan and execute complex software tasks while validating their own code. In addition to Gemini 3 Pro, Antigravity also integrates Gemini 2.5 Computer Use for browser control and Nano Banana (Gemini 2.5 Image) for image processing.
Safety and responsibility
Gemini 3 underwent the most extensive safety evaluations of any Google AI model to date. The model shows reduced sycophancy, increased resistance to prompt injections, and better protection against abuse via cyberattacks.
Google collaborated with international experts and provided early access to agencies such as UK AISI. Independent reviews came from Apollo, Vaultis, and Dreadnode. More information can be found in the Gemini 3 model card.
Gemini 3 is rolling out immediately for users of the Gemini app and for Google AI Pro and Ultra subscribers in AI Mode in Search. This is the first time Google has rolled out a Gemini model in Search on day one.
Developers can access it via the Gemini API in AI Studio, Google Antigravity, and Gemini CLI. Businesses can use the model via Vertex AI and Gemini Enterprise.
Google announces that more models in the Gemini 3 series will be released soon. The focus is on further developing intelligence, agents, and personalization to make AI truly useful for everyone.
Tip: Gemini 2.5 Pro available for paying GitHub Copilot users