Seven New Gemini Features Google Announced at I/O 2025

Google I/O's 2025 keynote could have more reasonably been called The Google AI Show. Almost everything the company talked about was AI-powered, some of which is promised to arrive in the future, and some of which is available today. Features were spread across Google's whole range of products, but here are some of the ones you're actually likely to see.

It's tough to talk about Gemini because it simultaneously refers to a set of models (like Gemini Flash, Gemini Pro, and Gemini Pro Deep Research), different versions of those models (the latest seems to be 2.5 for most of these), and different apps that these models are available through. There's the dedicated Gemini app, the voice assistant in things like Pixel phones and watches, as well as Gemini tools built into apps like Google Docs, Gmail, or Search.

I'll do my best to specify which features are coming to what products, but keep in mind that sometimes Google tends to announce the same thing a few times.

Agent Mode is coming to Gemini, Search, and more

The Gemini app is getting a new Agent Mode that can perform tasks for you while you do something else. Google showed off an example of asking Gemini to find apartments in a city. The app then searches listings online, filters them by the criteria you set, and can offer to set up apartment tours for you.

The most interesting aspect of this is that Google pitches this as a task you can have Gemini repeat regularly. So, for example, if you want Gemini to search for new apartments every week, the app can repeat the process, continuing with the information in previous iterations of the search.

Agent Mode is similarly coming to Google Search for certain requests. Google uses the example of asking for tickets to an upcoming event. Google scours ticket listing sites, cross-references against your preferences, and presents the results.

Gmail will pretend to be you when it replies to your emails

Gmail has had smart replies for a while, but they can sound pretty generic (without intervention, anyway). It's a dead giveaway to your recipient that you're not really paying attention. To help you get away with quietly ghosting your friends, Gmail will soon be able to tailor its responses to you by referring to your past emails and even Drive documents.

Google uses the example of a friend asking how you planned your recent vacation, a common thing we all email each other all the time. In this case, Gmail can draft a response based on your email history, with the advice you would be likely to give, and even write it how the AI thinks you would write it.

Thought summaries will summarize how AI summarizes its thought process

Yes, you read that right. AI "reasoning" models typically work by taking your query, generating text that breaks it down into smaller parts, sending those parts to the AI again, then carrying out each step. That's a lot of instructions happening behind the scenes on your behalf. Usually, reasoning models (including Gemini) will have a little drop down to show you the steps it took in the interim.

If even that is too much reading for you, Gemini will now summarize the summary of the thought process. In theory, this is to make it easier to understand why Gemini arrived at the answers it gives you.

Native audio output will whisper to you (in your nightmares)

This is technically a new feature of the Gemini API, which means developers can build on these tools in their apps. Native audio output will let developers generate natural-sounding speech. In its demo, Google showed off voices that could switch between multiple languages, which was pretty cool.

What isn't so cool, however, is the model can also whisper. I do not yet know what the practical use-cases are for an AI-generated voice that can whisper, but I do know I won't be able to get it out of my head for a week. At best.

Jules will fix your code's bugs in the background while you work

Last year, Google announced Jules, a coding agent that can help you with your code, similar to Github's Copilot. Now, the public beta of Jules is available. Google says Jules can fix bugs while you're working on other tasks, bump dependency versions, and even provide an audio summary of the changes that it's made to your code.

Google Search will let you virtually try on clothes while shopping online

I'm not great at visualizing what a piece of clothing will look like on my particular body, so this new try-on feature might actually be useful. Google is launching a Search Labs experiment that lets you upload a full-length photo of yourself that Google will alter to show what the clothing will look like on you.

The company is also integrating shopping tools that can buy items for you and even track for the best price. It will then be able to buy stuff for you via Google Pay, using your saved payment and shipping info. This one isn't available quite yet, and frankly we'd want to learn a little more about how the process works and how to prevent purchases you don't want before we'd recommend using it.

New Veo and Imagen models will generate audio and video

Video is, definitionally, a series of images played at a fast enough speed to convey a sense of motion. With that definition, I can confidently say that the demos of Google's new Veo 3 model does, in fact, show video. Whether that video is any good is in the eye of the beholder, I suppose.

Google seems to be betting on users finding the video generated by Veo 3 (and, by association, the images from Imagen 4) to be worthwhile, because the company is also building a video editing suite around it. Flow is a video editing tool that ostensibly lets editors extend and re-generate clips to get the right look.

Google also says that Veo 3 can generate sounds to go along with its video. For example, in the owl scene linked above, Veo also generates forest sound effects. We'll have to see how it generates these elements (can you edit individual sounds distinctly, for example?) but for now the demos speak for themselves. Veo 3 is now available in the Gemini app for Ultra subscribers.