Google I/O highlights agentic Gemini upgrades for search, shopping and video

Summary

Google used its Google I/O developers conference on Tuesday to push further into what it calls an “agentic” era, pitching Gemini updates that aim to help users complete tasks rather than only respond to questions. In remarks at a keynote near Google’s headquarters in Mountain View, California, CEO Sundar Pichai said the company is still working to make agents “easy to use, super secure and truly helpful,” while acknowledging that “it’s still early days” for the technology.

Google tied the push to what it said is broadening adoption of Gemini. Pichai said the Gemini app had 400 million monthly active users last year, and that usage has now surpassed 900 million, more than doubling over the period.

Alongside the agent pitch, Google detailed changes to its model lineup and where they will run first. The company said its latest family of models, Gemini 3.5, is rolling out Tuesday to billions of users beginning with Gemini 3.5 Flash, describing the Flash model as focused on speed and as its strongest agentic and coding model yet, while saying it is about four times faster than some competitors.

Google said Gemini 3.5 Flash will become the default model in the Gemini app and in “AI mode” on Google Search. The company also said it is working on Gemini 3.5 Pro, using it internally and expecting to launch next month, and it said Gemini 3.5 was developed with “new, more advanced safety training and mitigations,” aimed at reducing harmful outputs and reducing incorrect refusals for safe queries.

For media creation, Google announced Gemini Omni and said its tools will enable users to create high-quality video by providing a query with different kinds of inputs, including text, images, videos and audio. Google said videos generated with Omni can be edited through a conversation with the model, and it said the Omni family will eventually allow image and audio generation as well, while offering no timeline for those capabilities.

At the same time, Google described how it will try to authenticate AI-generated content. It said Omni Flash, the first Omni model, is launching Tuesday for Google’s “Al Plus, Pro and Ultra” subscribers in the Gemini app and Google Flow, and it will also be available at no cost on YouTube Shorts and YouTube Create App beginning this week. Google said videos created with Omni will include its SynthID “imperceptible digital watermark,” and it also said it will add “content credentials verification” to the Gemini app—designed to help determine whether content such as photos or video was AI-generated or captured with a phone camera and edited with AI tools—adding that the tool will appear in search in Chrome in coming months.

Google also used the conference to describe a new, always-on style of AI assistant. The company introduced Gemini Spark, saying it will complete routine tasks such as sorting meeting notes, emails and chats, and then creating a document with takeaways and to-do items. Google said Spark is based in the cloud, so it continues working in the background even when users shut their laptops or lock their phones.

Google said Spark’s proactive behavior would come with permission controls for some actions. The company said Gemini Spark is designed to ask for permission before performing “high-stakes” tasks like sending an email or making a purchase, and it said select testers will have access beginning Tuesday, with a plan to roll out a beta mode to U.S.-based subscribers to its Google AI Ultra tier. Later this summer, Google said Spark will operate directly within Chrome.

In search, Google’s updates build on “AI mode,” which it introduced last year and which it said lets users get more conversational answers before links. Google’s head of search, Liz Reid, said that AI mode queries have more than doubled every quarter since launch and recently surpassed 1 billion monthly users; for Tuesday’s update, Google said the default model in search will shift to Gemini 3.5 Flash.

Google also said it is introducing what it called an intelligent search box, describing it as the biggest upgrade to the search box in 25 years. Google said the box will adapt to accommodate longer queries and can help users write out their questions with AI-powered suggestions instead of traditional autocomplete, and it said users can search using multiple modalities, including text, images, video, files and even Chrome tabs as inputs. The company said the search box rollout begins Tuesday in all countries and languages where AI mode is already available.

For shopping, Google announced Universal Cart, which it described as an “intelligent shopping cart” that works across merchants and across services. The company said users can add items while browsing Google Search, chatting with Gemini, watching YouTube, or reading emails in Gmail, and that the cart would then run on Gemini models to look for deals and price drops, provide price history information and alert users when items come back in stock. Google said Universal Cart will be available this summer on search and the Gemini app, with YouTube and Gmail to follow.

Google also used the conference for consumer hardware updates. The company said it will offer two kinds of smart glasses, with audio glasses arriving first later this fall. Google said the audio glasses will let users say “Hey Google” or tap the side of the frame to access Gemini, which can assist with navigation, managing communications on their phone, real-time translations and other tasks.

Google said it partnered with Samsung and eyewear brands Gentle Monster and Warby Parker to create the glasses, showing two designs on Tuesday: sunglasses from Gentle Monster and glasses from Warby Parker. Google said those designs will launch as part of the eyewear brands’ full collections later this year.

Google’s overall message at I/O was that it is trying to make AI more actionable across products, from search and shopping to media creation and background task handling. Across those announcements, Google repeatedly pointed to guardrails such as permission requests for high-stakes actions and new watermarking and verification tools aimed at tracking AI-generated content.