Google released its AI model ‘Gemini 3.0’ which is the largest and most advanced of all. It continues to improve incrementally in real life use. It is also capable of working in almost all areas of advanced problem solving, reasoning, and in application development, especially agent use. It is especially recommended for coding, working with long contextual problems, and devising problem solving algorithms.

Gemini 3 is available now

At a press conference, Google executives emphasized that Gemini 3 leads several popular industry benchmarks for measuring AI model performance. Gemini 3 topped the global AI model LMARaena leaderboard with a record-breaking score of 1501, and achieved a top score of 37.5% in Humanity’s Last Exam benchmark, which measures general reasoning ability, surpassing the previous record of 31.64% held by GPT-5 Pro. The model also set new industry standards in multiple dimensions, including mathematics, multimodal understanding, and factual accuracy.

Demis Hassabis, CEO of Google’s DeepMind AI research lab, stated that Gemini 3 is “the world’s best multimodal understanding model” and the company’s most powerful agent and code generation model to date. Google CEO Sundar Pichai called it “our smartest model” in a blog post, capable of “turning any idea into reality.”

Google launched Gemini 3 just 11 months after the release of its second-generation Gemini model. This release coincided with Microsoft’s Ignite conference, only a week after OpenAI’s GPT 5.1 release, and two months after Anthropic’s Sonnet 4.5 release, highlighting the fierce competition in the development of cutting-edge AI models.

Gemini was designed from the outset to seamlessly integrate multimodal information on any subject, including text, images, video, audio, and code. Gemini 3 combines its advanced reasoning, vision, and spatial understanding capabilities, leading multilingual performance, and millions of context windows. In contrast, the maximum output of Claude Sonnet 4.5 and GPT 5.1 remained in the tens or hundreds of thousands range.

Gemini 3.0 has been launched on AI Studio, Gemini CLI, and other major developer portals such as Cursor, GitHub, JetBrains, and Cline.

Major Breakthrough in Reasoning Capabilities

Gemini 3 has made remarkable progress in reasoning capabilities, demonstrating PhD-level performance in multiple academic benchmark tests. The model achieved a high score of 91.9% on the GPQA Diamond test, set a record of 23.4% on the MathArena Apex benchmark in mathematics, and achieved a score of 72.1% in fact-accuracy on the SimpleQA Verified test. Gemini 3 also performed exceptionally well in multimodal reasoning, achieving 81% on the MMMU-Pro test and 87.6% on the Video-MMMU test. This means that the model can reliably handle complex problems across a wide range of fields, including science and mathematics.

Gemini 3 also performed exceptionally well in multimodal reasoning, achieving 81% on the MMMU-Pro test and 87.6% on the Video-MMMU test. This means that the model can reliably handle complex problems across a wide range of fields, including science and mathematics.

“We’ve seen a huge leap forward in inference capabilities with Gemini 3,” said Tulsee Doshi, Google’s product lead. “The depth and subtlety of its responses are things we’ve never seen before.”

In addition to the standard version, Google also launched the Gemini 3 Deep Think enhanced inference mode, which achieved a 41.0% score in the Humanity’s Last Exam test and a 93.8% score in the GPQA Diamond test. In the ARC-AGI-2 test, Deep Think mode achieved an unprecedented 45.1%, demonstrating its ability to solve novel challenges. Google stated that this mode is undergoing additional security evaluation and will be available to Google AI Ultra subscribers in the coming weeks.

Comprehensive Upgrade of Developer Tools

In the field of code generation, Gemini 3 is hailed by Google as “the best vibe coding and agent coding model ever built.” This model topped the WebDev Arena leaderboard with a score of 1487, achieved a score of 54.2% in the Terminal-Bench 2.0 test (measuring a model’s ability to operate a computer via a terminal), and reached 76.2% in the SWE-bench Verified benchmark (evaluating code agents), far surpassing the performance of Gemini 2.5 Pro.

Developers can access Gemini 3 through Google AI Studio, Vertex AI, Gemini CLI, and third-party platforms such as Cursor, GitHub, JetBrains, Manus, and Replit.

Google also launched Google Antigravity, a new agent-first development platform that leverages Gemini 3’s advanced inference, tool usage, and agent coding capabilities to transform AI assistance from a tool in the developer’s toolbox into an active partner. “Agents can work across your editor, terminal, and browser to help you build your applications in the best way,” said Koray Kavukcuoglu, CTO of DeepMind. While Google Antigravity’s core remains the familiar AI integrated development environment (IDE), its agents have been elevated to a dedicated interface and gained direct access to editors, terminals, and browsers. This allows agents to autonomously plan and execute complex end-to-end software tasks on behalf of users, while validating their own code. The platform also integrates tightly with Google’s latest Gemini 2.5 Computer Use browser control model and the top-tier image editing model Nano Banana.

Multimodal Understanding & Agent Capabilities in Tandem

Gemini 3 retains the cross-modal information synthesis capabilities inherent in the series from the beginning, seamlessly processing multimodal information such as text, images, video, audio, and code, and equipped with a context window of 1 million tokens.

Hassabis gives an example: if a user wants to learn traditional cooking, Gemini 3 can decipher and translate handwritten recipes in different languages, creating a shareable family cookbook. If a user wants to explore a new topic, it can provide academic papers, long video lectures, or tutorials, generating interactive flashcards, visualizations, or other formats to help the user master the ingredients. The model can even analyze videos of a user’s pickleball game, identify areas for improvement, and generate a training plan for overall motion improvement.

Regarding agent capabilities, Gemini 3 ranks first on the Vending-Bench 2 leaderboard, which tests long-term planning capabilities. The model maintains consistent tool use and decision-making abilities throughout a full year of simulated operation, achieving higher returns without deviating from the task. This means that Gemini 3 can represent users navigating more complex, multi-step workflows, such as booking local services or organizing inboxes, by combining deeper inference with improved tooling.

Multiple Core Products Live at Launch

Google adopted an aggressive product strategy, integrating Gemini 3 into multiple core products on its launch day. This marks the first time Google has applied a new model to its search product on the same day of its release, signifying a significant shift in the company’s AI commercialization strategy.

In search, AI Mode now uses Gemini 3 to deliver new generative user interface experiences, including immersive visual layouts, interactive tools, and simulations, all generated in real-time based on user queries. Gemini 3 is available to all users in the Gemini App, and Google AI Pro and Ultra subscribers can also use the model in AI Mode search.

For enterprise users, Gemini 3 is available in Vertex AI and Gemini Enterprise. Developers can access the model through the Gemini API, AI Studio, the Google Antigravity platform, and the Gemini CLI. Regarding security, Google states that Gemini 3 is its most secure model to date, having undergone the most comprehensive security assessment in the history of Google AI models. The model exhibits lower sycophancy, stronger resistance to cue injection, and better protection against abuse through cyberattacks. In addition to internal testing based on its cutting-edge security framework, Google collaborated with leading domain experts worldwide for evaluation, provided early access to organizations such as AISI in the UK, and received independent assessments from industry experts such as Apollo, Vaultis, and Dreadnode.