Today, Google has introduced Gemini 3, a significant enhancement to its premier multimodal model. The company claims that this new version excels in reasoning, features more seamless multimodal functionalities (the capability to operate across voice, text, or images), and will function as an agent.
The earlier model, Gemini 2.5, supports multimodal inputs. Users can provide it with images, handwriting, or speech. However, it typically needs explicit indications regarding the desired format in return, defaulting to plain text.
Gemini 3, however, brings forth what Google refers to as “generative interfaces,” enabling the model to autonomously determine the most suitable type of output for the prompt, creating visual arrangements and dynamic displays independently instead of merely providing a block of text.
When requesting travel suggestions, it might generate a website-like interface within the application, complete with sections, images, and follow-up inquiries like “How many days are you traveling?” or “What types of activities do you prefer?” It also offers clickable choices based on your potential next steps.
When asked to clarify a concept, Gemini 3 may draw a diagram or create a basic animation if it deems a visual representation more effective.
“Visual layout crafts an engaging, magazine-style presentation complete with images and sections,” states Josh Woodward, VP of Google Labs, Gemini, and AI Studio. “These components not only look appealing but also encourage your input to refine the results further.”
With the launch of Gemini 3, Google is also introducing Gemini Agent, a novel feature intended to tackle multi-step activities directly within the application. The agent can link to services like Google Calendar, Gmail, and Reminders. Upon receiving authorization, it can perform tasks such as organizing an inbox or coordinating timelines.
Similar to other digital agents, it divides tasks into distinct steps, shows its progress in real-time, and awaits user approval before moving on. Google describes this feature as a move toward “a genuine generalist agent.” It will be accessible on the web for Google AI Ultra subscribers in the US starting November 18.
This approach can resemble “vibe coding,” where users articulate an end objective in straightforward language and let the model construct the required interface or code to achieve it.
The update also integrates Gemini more closely with Google’s existing services. In Search, a limited number of Google AI Pro and Ultra subscribers can now shift to Gemini 3 Pro, the reasoning variant of the updated model, to acquire deeper, more comprehensive AI-generated summaries that depend on the model’s reasoning rather than the former AI Mode.
For shopping, Gemini will now draw from Google’s Shopping Graph—which the organization claims features over 50 billion product listings—to generate its own recommendation guides. Users merely need to pose a shopping-related query or input a shopping-related term, and the model will compile an interactive, Wirecutter-style product recommendation document, complete with pricing and product specifics, without redirecting to an outside site.
For developers, Google is also advancing single-prompt software creation. The company unveiled Google Antigravity, a development platform that serves as an all-encompassing space where code, tools, and workflows can be produced and managed from a single command.
Derek Nee, CEO of Flowith, an agentic AI application, informed MIT Technology Review that Gemini 3 Pro addresses multiple shortcomings in previous iterations. Enhancements include improved visual comprehension, superior code generation, and enhanced performance on extended tasks—traits he considers vital for creators of AI applications and agents.
“Due to its speed and cost benefits, we’re integrating the new model into our offering,” he states. “We’re hopeful about its potential, but more extensive testing is essential to assess its full capabilities.”
