Google I/O 2025: Sundar Pichai Keynote Highlights
Source: blog.google
Here's how AI is becoming more helpful with Gemini. Below is an edited transcript of Google CEO Sundar Pichai’s remarks at Google I/O 2025, adapted to include more of what was announced on stage.
Normally, there would not be news in the weeks before I/O, because the best models were saved for the stage. Now, in the Gemini era, the most intelligent model might ship on a Tuesday, or a breakthrough like AlphaEvolve might be announced a week before. The goal is to get the best models into the hands and products ASAP, and so shipping is happening faster than ever.
Rapid Model Progress
Model progress is exciting. Elo scores are up more than 300 points since the first-generation Gemini Pro model. Gemini 2.5 Pro sweeps the LMArena leaderboard in all categories. This model progress is enabled by infrastructure. The seventh-generation TPU, Ironwood, is designed to power thinking and inferential AI workloads at scale. It delivers 10 times the performance over the previous generation, and packs 42.5 exaflops compute per pod.
This infrastructure helps deliver faster models, even as model prices decrease. Google is leading the Pareto Frontier, and fundamentally shifted the frontier itself. More intelligence is available, for everyone, everywhere. The world is adopting AI faster than ever. All this means that we’re in a new phase of the AI platform shift, where decades of research are becoming reality for people, businesses and communities.
New Platforms and Features
Project Starline, a 3D video technology, debuted at I/O a few years back, to create a feeling of being in the same room as someone far away. Technical advances have continued and now Google Beam, a new AI-first video communications platform is being introduced. Beam uses a state-of-the-art video model to transform 2D video streams into a 3D experience, using six cameras and AI to merge video streams and render you on a 3D lightfield display. It has near perfect head tracking, and runs at 60 frames per second in real-time. The first Google Beam devices will be available for early customers later this year, in collaboration with HP.
More immersive experiences have also been created in Google Meet, including speech translation to break down language barriers. In near real time, it can match the speaker’s voice and tone, and expressions. Translation in English and Spanish is rolling out to Google AI Pro and Ultra subscribers in beta, with more languages coming soon. This will come to Workspace business customers for early testing this year.
Project Astra, first seen at I/O, explores the future capabilities of a universal AI assistant. Gemini Live now incorporates Project Astra's camera and screen-sharing capabilities, already available to all Android users and rolling out to iOS users. These capabilities are coming to products like Search.
Agents are systems that combine the intelligence of advanced AI models with access to tools, to take actions on your behalf and under your control. Project Mariner is an early research prototype in agents with computer-use capabilities to interact with the web and get stuff done. It was released as an early research prototype, and progress has been made with new multitasking capabilities and a method called “teach and repeat”. Project Mariner’s computer use capabilities are coming to developers via the Gemini API. Trusted testers are already starting to build with it, and it will be available more broadly this summer. The Gemini API and SDK are now compatible with MCP tools.
Agentic capabilities are also starting to come to Chrome, Search and in the Gemini app. A new Agent Mode in the Gemini app will help you get even more done. An experimental version of Agent Mode in the Gemini app will be coming soon to subscribers.
Personalization and Search
Personalization will be powerful. Gemini models can use relevant personal context across your Google apps in a way that is private, transparent and fully under your control. One example of this is personalized Smart Replies in Gmail. If your friend emails you for advice about a road trip that you’ve done in the past, Gemini can do the work of searching your past emails and files in Google Drive, such as itineraries you created in Google Docs, to suggest a response with specific details that are on point. It will match your typical greeting and capture your tone, style and even favorite word choices, all to generate a reply that’s more relevant and sounds authentically like you. Personalized Smart Replies will be available for subscribers later this year.
Gemini models are helping to make Google Search more intelligent, agentic and personalized. AI Overviews have scaled to over 1.5 billion users and are now in many countries and territories. As people use AI Overviews, they’re happier with their results, and they search more often. For those who want an end-to-end AI Search experience, an all-new AI Mode is being introduced as a total reimagining of Search. With more advanced reasoning, you can ask AI Mode longer and more complex queries, and you can go further with follow-up questions. All of this is available as a new tab right in Search. AI Mode is coming to everyone in the U.S.. With the latest Gemini models AI responses are fast and accurate. Gemini 2.5, is coming to Search in the U.S., as well.
Gemini 2.5 Flash has been incredibly popular with developers who love its speed and low cost. The new 2.5 Flash is better in nearly every dimension. An enhanced reasoning mode called Deep Think is being introduced to make 2.5 Pro even better. Deep Research is becoming more personal, allowing you to upload your own files and soon connect to Google Drive and Gmail, enhancing its ability to generate custom research reports. It's also being integrated with Canvas, enabling the creation of dynamic infographics, quizzes and even podcasts in numerous languages with a single click. Camera and screen sharing capabilities are becoming freely available to everyone for Gemini Live, including iOS users, and will soon connect to Google apps for more seamless assistance.
Veo 3, a state-of-the-art video model, now has native audio generation. Imagen 4, the latest and most capable image generation model, is also being introduced. Both are available in the Gemini app. Filmmakers can use a new tool called Flow to create cinematic clips, and extend a short clip into a longer scene.
Research is being done today that will become the foundation of tomorrow’s reality, from robotics to quantum, AlphaFold and Waymo. A recent experience highlighted this. While in San Francisco with my parents, they wanted to ride in a Waymo. My father was amazed, and it was a reminder of the power of technology to inspire.