Google LLC has come up with the perfect response to the bevy of artificial intelligence announcements at Microsoft Ignite this week, launching its most intelligent model: Gemini 3.

The launch of Gemini 3 has been hotly anticipated for several months, and Google is doing everything it can to ensure that users won’t be disappointed. As the company’s smartest model yet, Google says, it possesses industry-leading reasoning capabilities and the ability to learn almost any new concept. It’s being made broadly accessible from the get-go, launching in the Gemini application, Google Search and developer platforms such as AI Studio, Vertex AI and Google Antigravity – an all new agentic development platform.

Smarter, more thoughtful

In a blog post, Google DeepMind Chief Executive Demis Hassabis claimed that Gemini 3 is the “best model in the world for multimodal understanding,” as well as one of the most capable vibe coding models ever released. This is based on its high-level reasoning capabilities, which enable Gemini 3 to outperform its predecessor Gemini 2.5 on every major AI benchmark.

Gemini 3 has jumped straight to the top of the LMArena Leaderboard with an all-time high score of 1,501 points, Hassabis said, and it also demonstrates Ph.D.-level reasoning capabilities with a 37.5% score on Humanity’s Last Exam and a 91.9% score on GPQA Diamond. Moreover, it excels at mathematics, achieving a new high score of 23.4% on the MathArena Apex benchmark, while its factual accuracy also sets a new standard for LLMs, with a 72.1% score on SimpleQA Verified.

Hassabis said users will immediately notice that Gemini 3 brings a lot more nuance and depth to every interaction it has with them, with smarter, more concise and direct responses, “trading cliche and flattery for genuine insight”. Google has worked hard to ensure Gemini 3 tells users what they need to hear, rather than what they want to hear, Hassabis explained, allowing it to act more like a “true thought partner”.

And if Gemini 3 doesn’t prove smart enough when pressed with more complex queries, users will be able to activate a new “Deep Think” mode, which enhances its reasoning skills by giving it more time to ponder before it generates a response.

In addition, Gemini 3 boasts a unique ability to synthesize information about any topic across multiple modalities at once, including text, video, images, code and audio. “For example if you want to learn how to cook in your family tradition, Gemini 3 can decipher and translate handwritten recipes in different languages into a shareable family cookbook,” Hassabis said.

Alternatively, if people wants to educate themselves about any new topic they’re unfamiliar with, they can feed Gemini 3 with various academic papers, video-based lectures and tutorials, and it will come up with a concise lesson, including interactive flashcards and visualizations, to help the user digest the material more easily.

In addition, anyone who explores the AI Mode in Search will be treated to new Gemini 3-enabled experiences, such as immersive visual layouts, interactive tools and simulations, which will be quickly generated based on their query.

Constellation Research analyst Holger Mueller said Gemini 3’s impressive benchmarks speak for themselves, and underscore his long-held view that Google is well ahead of its competitors in areas such as multimodal AI. “Gemini 3 shows that, with users and tasks being able to seamlessly transition between different modalities,” he said. “It’s innovating fast, and by making Gemini 3 available immediately in Search, Vertex AI and other platforms, it demonstrates Google’s strong confidence in its capabilities.”

Higher-level agentic coding

Developers will likely be among the most avid users of Gemini 3, for Hassabis claims it’s now the industry’s best vibe coding model, thanks to its ability to handle more complex prompts and instructions. His claim is backed up by yet more benchmarks, with the model achieving a score of 1,487 on ELO and 54.2% on Terminal-Bench 2.0, which tests models for their ability to use a computer via the terminal.

Hassabis said developers will be able to get started writing code with Gemini 3 from today in AI Studio, Vertex AI and the Gemini Command Line Interface tool, as well as third-party developer tools such as Cursor, GitHub, JetBrains, Replit and Manus. However, the most exciting way to access Gemini 3 may be with Google Antigravity, which is a new agentic development platform that supports agentic automation at a much higher, task-oriented level.

Antigravity promises to transform developer productivity by evolving AI from an assistant into a true partner that’s able to perform work on its own initiative. The platform looks much like a typical coding environment, but the agents have been granted direct access to the editor, the terminal and a web browser, giving them all of the tools they need to autonomously plan and execute complex programming tasks on behalf of humans.

The Antigravity platform is powered by Gemini 3, and it’s also tightly integrated with Google’s Gemini 2.5 Computer Use model that provides advanced browser control, as well as its best image editing model, Nano Banana. “With Antigravity, the agents are acting much more autonomously than before, and it’s yet more evidence of Google’s continued leadership in terms of multimodal reasoning and AI coding,” Mueller said.

Stronger security

As always with new AI models, there’s a security angle, and Google claims that Gemini 3 is also the most secure it has yet released. It said the model has undergone the most extensive set of evaluations of any of its models, being tested against critical domains in its own Frontier Safety Framework, and also by third-party organizations such as Apollo, Vaultis and Dreadnode.

The results of those tests show that Gemini 3 demonstrates significantly reduced sycophancy and higher resistance to prompt injection attacks and malware-based attacks, Hassabis said. Meanwhile, Gemini 3’s Deep Think mode is currently undergoing additional safety evaluations and input safety tests, prior to being made available to Google AI Ultra subscribers in the coming weeks.

Images: Google