Google I/O 2025: Gemini, AI Agents, Generative Media and XR

I just returned from Google I/O 2025, where Google showed a wide range of AI products. I also interviewed CEO Sundar Pichai about world models, search and an intelligence explosion, and plan to share that conversation separately.

A year earlier, many people doubted Google’s AI strategy. Since then, it has announced or released AlphaFold 3, Imagen 3, Gemma 2, Project Mariner, Gemini 2.5, Gemma 3 and AlphaEvolve. My impression of this event is that Google is moving research into products at a rapid pace.

One striking company metric is monthly tokens processed across Google products and APIs: 9.7 trillion a year earlier versus more than 480 trillion in 2025, roughly 50 times as many. That is Google’s aggregate measure, not a direct measurement of individual users or time saved. Model usage, product integration and longer reasoning outputs can all contribute.

Project Starline has become Google Beam, a three-dimensional video communications platform. Pichai says the goal is to make distant participants feel as though they are in the same room. Google showed it as an enterprise-oriented system rather than an ordinary consumer screen.

I tried Beam in person. Multiple cameras capture the other person and the display produces a depth effect reminiscent of a Nintendo 3DS. My eyes needed a moment to adjust; when someone held up an apple, it felt close enough to reach. That is my impression of a demonstration, not a test of every deployment.

Project Astra’s capabilities are also reaching Gemini Live: a phone camera can let users converse about what they are seeing and ask questions about the scene. Google’s comic demonstration has a user mistake a garbage truck for a convertible, a streetlight for a building and a shadow for a follower. Camera and screen-sharing rollout details varied by platform and region.

Project Mariner is Google’s web-interaction agent. The announced multitasking feature points toward a workflow in which an agent handles one task while a user starts another; computer use, tools and memory come together. These are early systems, and a demonstration is not proof that long tasks reliably succeed without oversight.

Google also described agent capabilities for Chrome, Search and the Gemini app. In a keynote example, Pichai asks for an Austin apartment for three roommates with a $1,200 monthly budget each and laundry access. The planned agent searches listings, uses Mariner for filters and can use MCP-connected tools to arrange a tour; this is a demonstration of a proposed workflow.

What excites me most is a personal assistant that can use context from services such as Gmail, Calendar and YouTube, alongside memory from earlier interactions. I would like to open my inbox and find draft replies tailored to my history with each contact, ready for me to review and send. Personalized smart replies move partway toward that vision, but are not the full workflow I imagine.

I was also pleased that Google referenced my work with Gemini and a Rubik’s Cube demo during the presentation. The company introduced model updates including thinking budgets, faster performance and summaries. Thinking budgets offer a way to trade off reasoning-token use, latency and cost in supported configurations.

Google introduced Gemini Diffusion, a research model that iteratively refines text or code from noise. I find the speed of the demonstration striking, though model quality and latency depend on the specific system and measurement; diffusion is not automatically faster or less accurate than every autoregressive transformer. Google presented it as research, not a broadly proven replacement.

Gemini 2.5 Pro Deep Think was introduced as an experimental enhanced-reasoning mode using techniques including parallel approaches. In Google’s keynote chart, the company reported 49.4% on USAMO 2025, 80.4% on LiveCodeBench v6 and 84.0% on MMMU, a multimodal benchmark. These are reported scores for specific setups and competitors, not a universal ranking; the third benchmark is MMMU, not MMLU.

Google also spoke about moving toward world models. Demis Hassabis points to representations of nature, Veo’s video examples and Gemini Robotics as steps toward systems that can handle the physical world. The examples are demonstrations and the broader goal remains a research ambition, not proof of comprehensive physical understanding.

Imagen 4 generated detailed examples including a dress, paper-style bird, flowers and a cat. Google announced an Imagen 4 Fast version planned to be up to ten times faster than Imagen 3; the figure is a company claim about that specific forthcoming variant, not a measured speedup for every Imagen 4 request.

Veo 3 generates video with audio, including dialogue and sound effects, as shown in clips of animals and the ocean. Google also introduced the music model Lyria 2. I plan to test the tools myself; polished examples selected for a keynote do not independently establish typical output quality.

At the event Google AI Ultra was announced in the US at $249.99 per month, with a first-time three-month discount and higher limits and early access to features. I consider it expensive but plan to subscribe and report back. That is the 2025 launch price and availability, not a current price quote.

Flow combines Google’s generative media models into a filmmaking tool. In the demo, a creator generates ingredients such as a chicken-shaped gearshift, supplies a camera-angle prompt and combines images to create a clip. The interface can arrange and extend clips, giving creators more control than a single prompt alone.

The resulting sequence combines generated visuals and sound. I compare Flow with Sora’s scene-building tools and express my preference for Veo 3’s video quality; that comparison is my judgment, not an independent benchmark. I intend to do a more extensive test later.

Finally, Google showed prototype Android XR glasses with information displayed in the lens. A live demonstration showed a backstage wearer receiving a message, checking the weather and looking at navigation. There was some visible jitter; I suspect network conditions, but the video does not establish the cause.

The demo shows a maps recommendation and directions over the wearer’s view. I have been skeptical that glasses are the ultimate AI form factor: I would wear them outdoors but do not want to wear them all day indoors. The prototype is a demonstration, not a guarantee of a released product or everyday performance.