A wide-angle shot of the number 10 in orbit around a space station.
Generated with Imagen and Veo.
This place is going to be mind-blowing. Generated with Veo 3.
[Music]
Wake up, kid, we're deep in a daydream. Fourteen years old, always bowing to the world, but still being a polite, good kid.
Music is stirring in your soul. Don't give up, music is stirring in your soul.
The last dance is about to start. The world will get better. Don't give up, you have a reason to live.
Don't forget, you reap what you sow.
Welcome to Google I/O.
I'm going home. You're the best, show me now.
Great video. It looks like the team had a lot of fun making it. Good morning, everyone. Welcome to Google I/O.
Welcome to everyone here at Shoreline and to everyone watching online around the world. Someone told me that Gemini season starts today. I don't think that's particularly special, because for Google, every day is Gemini season. In the weeks leading up to I/O, we usually don't have much activity, because we want to save our best models for this stage. But in the Gemini era, we might also release our most intelligent model on a random Tuesday in March, or a remarkable breakthrough like AlphaEvolve a week before the conference. We just want to get our best models into your hands as quickly as possible so you can experience our products. To that end, we are releasing far more frequently than ever before.
Since the last I/O, we have released over a dozen models and research breakthroughs, and over 20 major AI products and features. I am thrilled by the rapid progress of our models. You can see the step-change improvement in this chart. The Elo score is a measure of progress. From the first generation Gemini Pro, this score has improved by over 300 points. Today, Gemini 2.5 Pro dominates all categories on the LMArena leaderboard.
It excels in many benchmarks. In coding, we are also making rapid progress. The updated 2.5 Pro leads the WebDev Arena, with an Elo score 142 points higher than the previous version. This model is highly sought after on major coding platforms. Thanks to everyone's support. In the leading AI code editor Cursor, Gemini has become the fastest-growing model this year, with hundreds of thousands of lines of code generated with it every minute.
This last milestone might be the most touching, at least in some circles. A few weeks ago, Gemini successfully completed Pokémon Blue. It collected all eight badges, made it to Victory Road, and defeated the Elite Four and the Champion. This brings us one step closer to achieving API, though we are talking about 'Artificial Pokémon Intelligence'.
All of this progress is supported by our exceptional infrastructure, which is the cornerstone of our full-stack AI approach. Our seventh-generation TPU, Ironwood, is designed with cutting-edge technology to empower large-scale thinking and reasoning. Its performance is ten times that of the previous generation, delivering 42.5 exaFLOPS of computing power per pod. It's impressive. Google Cloud customers will be able to experience its power later this year.
Leveraging our infrastructure strength, starting from the underlying TPU, we offer models with superior speed. Gemini holds the top three spots on the LMArena leaderboard, achieving the highest tokens per second and significantly reduced model prices. Balancing price and performance is not easy, but we have done it time and again, providing excellent models at the most effective price point. Google is not only ahead on this Pareto frontier but has fundamentally changed it, making better intelligence accessible to everyone, anytime, anywhere.
Our efforts have received a positive response. AI adoption is faster than ever. As a sign of progress, this time last year, our total monthly token processing across all products and APIs was 9.7 trillion. Today, it is 480 trillion tokens per month. That's nearly a 50-fold increase in just one year.
The adoption of our developer AI tools has become a wave. In Google AI Studio and Vertex AI, over 7 million developers are using the Gemini API. That's more than five times the number at last year's I/O. Since last year, Gemini usage in Vertex AI has increased 40 times.
Our products are also increasingly adopting AI. The monthly active users of the Gemini app have exceeded 400 million. Growth and engagement are strong, especially with the 2.5 models. For users of the 2.5 Pro model in the Gemini app, usage has increased by 45%. You'll hear more about the Gemini app later. In Search, we are also seeing tremendous momentum. AI Overviews now have over 1.5 billion monthly users.
This means that the number of people using generative AI through Google Search exceeds that of any other product globally. Beyond AI Overviews, AI Mode is another major advancement in Search. We'll share more about that later. All of this progress points to one thing: we are in a new phase of the AI platform shift, where decades of research are being transformed into practical tools that benefit everyone.
I want to share three examples of how research is fundamentally changing our products today: Project Starline, Astra, and Mariner. At I/O a few years ago, we unveiled Project Starline, our breakthrough 3D video technology that makes you feel like you're in the same room with someone far away. We've been improving the technology, and today, we are finally ready to start its next chapter. Introducing Google Beam, a new AI-first video communication platform.
Beam uses a new advanced video model to transform 2D video streams into a lifelike 3D experience. Behind the scenes, a six-camera array captures your image from different angles. Then, AI merges these video streams to generate a 3D light field, showing your stereoscopic image. It achieves millimeter-precision head tracking and real-time rendering at 60 frames per second, creating an incredibly natural, immersive conversation experience.
We are eager to bring this technology to everyone and have partnered with HP to do so. The first Google Beam device will be available to initial customers later this year. HP will share more details in the coming weeks, so stay tuned.
For years, we have been working hard to integrate the underlying technology of Starline into Google Meet. This includes real-time voice translation to help break down language barriers. This example shows its practical value when booking a vacation rental in South America without speaking the local language. Let's take a look. Hi, Camilla. Let me turn on the voice translation feature. It's great to finally talk to you directly.
[Spanish translation]
[Spanish response] You'll be happy. I'm sure you'll love this city too. The neighborhood this room is in is nice, and it overlooks the mountains. That sounds great.
[Spanish translation]
[Spanish response] There's a bus nearby, but I recommend renting a car to make it easier to explore the outdoors. Sounds good. [Spanish translation]
As you can see, the voice translation vividly mimics the speaker's tone, cadence, and even expressions. We are one step closer to natural, seamless cross-language communication. We have now integrated this real-time voice translation directly into Google Meet. Subscribers can use the English-to-Spanish translation starting now, and we will add more languages in the coming weeks. The real-time translation feature will be available to enterprise users later this year.
Another early research project that debuted on the I/O stage is Project Astra, which explores the future ability of a general-purpose AI assistant to understand the world around you. We are integrating it into Google products. Gemini Live now integrates Project Astra's camera and screen-sharing capabilities, allowing you to have conversations about anything you see. People are using it for various purposes, from practicing for job interviews to training for a marathon. We always highly value feedback from trusted testers, and also some less 'trusted' ones. Let's take a look.
What a beautiful convertible. I think you mistook a garbage truck for a convertible. What else can I help you with? What is this tall, thin building doing in front of my house? That's a streetlamp, not a building. Why are these palm trees so short? It's a bit worrying. They're not short; they're actually very tall. A handsome convertible. Still a garbage truck. Anything else? Why do people always put packages on my lawn? That's not a package; it's a utility box. Why is this person following me when I walk? No one is following you; that's just your shadow.
Gemini is really good at pointing out your mistakes. Starting today, this feature is available to all Android and iOS users.
Let's talk about our research prototype, Project Mariner. This agent can interact with the web and complete various tasks. We think outside the box, viewing the agent as a system that combines the intelligence of advanced AI models with the ability to use tools. They can autonomously perform actions for you, all under your control. Using a computer is an important capability for an agent, allowing it to interact with browsers and other software and perform actions on them.
Project Mariner is an initial attempt to test computer use capabilities. We released it as an early research prototype last December, and it has made significant progress since then. First, we introduced multitasking. It can now manage up to 10 tasks simultaneously. Second, it has a feature called 'Show and Repeat'. You only need to show it a task once, and it can learn to plan for similar tasks in the future. We will provide developers with Project Mariner's computer use capabilities through the Gemini API. Trusted testers like Automation Anywhere and UiPath have already started building products with it. We will make it available to a wider audience this summer.
To foster a thriving agent ecosystem, we need to develop a broader set of tools. Computer use capability is one part of this effort. Examples include our open Agent-to-Agent Protocol, which allows agents to talk to each other. We released this protocol at Cloud Next and have received support from over 60 technology partners, a number that continues to grow. There's also the Model Context Protocol from Anthropic, which allows agents to access other services. Today, we are thrilled to announce that the Gemini SDK is now compatible with MCP tools.
These technologies will work together to make agents more useful. We have already started adding agent capabilities to Chrome, Google Search, and the Gemini app. Let me show you the feature in the Gemini app that excites us the most. We call it Agent Mode. Imagine you want to find an apartment in Austin for yourself and two roommates. Each of you has a monthly budget of $1,200, and you want an apartment with a washer and dryer, or at least a laundromat nearby.
Normally, you'd spend a lot of time browsing countless listings. But with Agent Mode, the Gemini app can work for you. It accesses sites like Zillow in the background to find listings that match your criteria, using Project Mariner to adjust very detailed filters when needed. If there's a specific apartment you want to know more about, Gemini uses MCP to access the listing and can even help you schedule a viewing. It will keep browsing for new listings as long as you need, freeing up your time to do what you want, like planning a housewarming party. This is particularly useful for companies like Zillow, bringing in new customers and improving conversion rates.
An experimental version of Agent Mode in the Gemini app will be available to subscribers soon. This is an emerging new field, and we are excited to explore how to bring the benefits of agents to users and the broader ecosystem.
The best way to bring research results to life is to ensure they have real practical value in your actual environment. This is where personalization comes in. To that end, we are introducing a personal context feature. With your permission, the Gemini model can use relevant context across your Google apps in a privacy-protecting, highly transparent way that is fully under your control. Let me show an example in Gmail.
You might be familiar with our AI-powered Smart Reply feature, which is surprisingly popular. Wouldn't it be great if the AI-generated replies could mimic your tone? That's the idea behind personalized Smart Reply. Imagine a friend writes to me asking for advice. He's planning a road trip to Utah and remembers I've driven there before. Honestly, if I wrote the reply myself, I'd just give a few brief sentences that wouldn't be very helpful. Sorry, Felix.
But with personalized Smart Reply, I can be a helpful friend because Gemini does almost all the 'homework' for me. It looks at notes in my Drive, checks booking info in old emails, and finds itineraries in Google Docs. For the Zion National Park trip, Gemini mimics the greetings from my past emails, capturing my tone, style, and word preferences. It then automatically generates a reply. What I appreciate most is the thoughtful details it adds, like keeping driving time under five hours a day and using my favorite adjective: 'thrilling'.
Looks good. Maybe a few tweaks, and I can hit send. This feature will be available to subscribers in Gmail this summer.
You can imagine how useful the personal context feature would be across Search, Docs, Gemini, and other products. Today, I've talked about intelligence, agents, and personalization. These are the frontiers we will continue to push. Throughout the keynote, you'll hear more examples. But first, I'd like to invite someone to share more details about the intelligence driving our future innovations.
Last year, when I introduced him, I called him Sir Demis. This year, we can also add 'Nobel laureate' to his title. Demis, please come up.
Hello everyone! It's great to be back at Google I/O. We are at an important historical moment. AI is unlocking an exciting new future. This year, we have explored tirelessly and reaped abundant rewards. As Sundar said, people love interacting with Gemini 2.5. We've seen you use it for vibe coding, developing apps and games in one go, and using its powerful reasoning abilities for various tasks, from dissecting scientific papers to parsing YouTube videos. You've told us Gemini is collaborative, insightful, and a true assistant. We are thrilled to see the amazing things you're creating with it.
Gemini 2.5 is our most intelligent model to date and a world-class foundation model. Just two weeks ago, we released an upgraded preview of 2.5 Pro so you could start turning your ideas into reality. Your results have been eye-opening, from turning sketches into interactive apps to simulating 3D cities. The new 2.5 Pro model leads the popular coding leaderboard, WebDev Arena. Now, it also integrates LearnLM, a series of models developed with support from education experts, making 2.5 Pro the world's best learning assistant model, topping all leaderboards on LMArena.
Gemini Flash is our highly efficient flagship model, beloved by developers for its excellent speed and low cost. Today, I'm excited to announce the updated 2.5 Flash. The new Flash model has improved in almost every aspect, achieving better scores on all important benchmarks, including reasoning, coding, and long context. In fact, it's second only to 2.5 Pro on the LMArena leaderboard. I'm happy to say Flash will officially launch in early June, with Pro following shortly after.
We are making final adjustments based on your feedback, but you can try the preview now in AI Studio, Vertex AI, and the Gemini app. As always, we will continue to push technology forward. We are very pleased with our recent research breakthroughs. I'll ask Tulsee to share more later. But first, let's look at what you can do with Gemini today.
[30 things you can do with Gemini: Physics simulation, nature simulation, optical simulation, puzzle simulation, turning images into code, turning ideas into 3D models, adding 3D effects to paintings, co-creating with AI, hairstyle design, dressing up this cat, searching through decades of files, recipes, searching through hours of video, building robots, coding just by talking. Can you make a confetti cannon effect coming out of my mouth? I updated the sketch. AI for everything: entertainment, learning, animation, visualization, creation, transcription, practice, translation, query, emoji, remix, calculation, exploration, understanding, maps, explanation. Generate a thousand things, imagine ten thousand possibilities. What should we do next? Here are 30 more ideas for you. Start building with Gemini. AI Studio.
Demis, thank you! I'm glad to share that to help developers like you build more easily with Gemini 2.5, we've made several improvements based on your feedback. These include feature improvements, enhanced security and transparency, improved cost efficiency, and better controllability. First, in addition to the new 2.5 Flash that Demis mentioned, we are also launching a new preview of text-to-speech. Based on native audio output, it offers industry-first multi-speaker support, capable of handling two voices. This means the model can engage in more expressive conversations, capturing the nuances of our speech and even seamlessly switching to a whisper, like this.
It supports over 24 languages and can even switch between languages fluently. The model can start in English, then [Hindi], and switch back, with the voice remaining consistent. Pretty cool, right? Starting today, you can use this text-to-speech feature in the Gemini API. Later today, the Live API will also offer a preview of native audio conversation with 2.5 Flash, allowing you to converse with Gemini more naturally. It can even distinguish between speakers and background noise to determine the right time to respond. Logan and Paige will share more in the developer keynote.
Second, we have strengthened protection against security threats like indirect prompt injection, making Gemini 2.5 our safest model yet. In both the 2.5 Pro and Flash models, we have added a thinking summary feature, available through the Gemini API and Vertex AI. The thinking summary extracts the model's raw thought process and organizes it into a clear, structured format with headings, important details, and information about model actions like tool calls. This gives you clearer insight into the model's thinking, which is particularly useful for high-latency tasks and for debugging and usage.
Everyone has also called for improved model efficiency, and now 2.5 Flash is more efficient than ever. It uses fewer tokens for the same performance, achieving a 22% efficiency improvement in real-world tests. Finally, we have introduced a thinking budget for 2.5 Flash, allowing you to balance cost, latency, and output quality. Due to great feedback, we have also added the thinking budget to 2.5 Pro, which will be released with our official model in the coming weeks. With the thinking budget, you can better control how many tokens the model uses to think before answering, or you can turn thinking off entirely.
As Demis mentioned, Gemini 2.5 Pro has exceptional coding abilities. So next, I want to show how to turn an idea into reality. If you've visited the American Museum of Natural History in New York City, you'll remember its wonderful exhibits. To share these exhibits, I had 2.5 Pro in Google AI Studio write a simple web app to share some photos and detailed information. This is what the app looks like so far. But I want to make it more interactive. I'm still brainstorming design ideas, but I have some thoughts. You've seen situations like this where someone finds you and shows you a 'brilliant' idea sketched on a napkin. As a PM, I'm often that person with the napkin.
Regular 2D web design is already challenging, but I want to go 3D. I know moving to 3D isn't easy; there's a lot to learn, like setting up scenes, cameras, and lighting. But luckily, 2.5 Pro can help me. What I'm going to do is add the sphere image I just showed you, input a prompt for 2.5 Pro to update my code based on this image. We let 2.5 Pro start working. You can see it begins to think and creates a plan based on my request, which it will then apply to my original codebase.
Gemini is a multimodal model, so it can understand this abstract sphere sketch and generate beautiful 3D animations through code, then apply it to my original app. This process takes about two minutes. To save time, we'll do a 'bake-off' style presentation. I'll switch to another tab. Before this keynote started, I ran the same prompt. This is the result Gemini generated.
We turned a rough sketch directly into code, updating multiple files. You can see it thought for 37 seconds, and you can see its thought process and which files it updated. All of this was done in AI Studio. So after prototyping, I can deploy the code directly with my Gemini API key. This is the final app in Chrome. Look at these animations. I didn't need deep knowledge of the Three.js library or complex 3D algorithms to create this. If I had to do it myself, I might never have achieved it, but here I did it with just a sketch.
I can use multimodality to enrich the entire experience. I used 2.5 Flash to add a question to each photo, inviting you to learn more. But what if it could talk? That's where we can use Gemini's native audio capabilities. This is a pangolin. Its scales are made of keratin, just like your fingernails. Wow! Now it has voice capabilities. You can hear how to add expressive audio to your app. Before sharing more, I want to end the demo with another fun design, also coded with 2.5 Pro.
We have provided the starting code in Google AI Studio so everyone can build upon this demo. Feel free to try it. This demo is just one example proving that Gemini 2.5 will fundamentally change how we build, and how you use its extraordinary capabilities to integrate vision, voice, and coding into your own apps. You can use 2.5 Pro in your favorite IDE platforms and various Google products, such as Android Studio, Firebase Studio, Gemini Code Assist, and our latest asynchronous coding agent, Jules.
You just submit a task, and Jules handles the rest: fixing bugs, performing updates. It integrates with GitHub and can run independently. Jules can handle complex tasks in large codebases that used to take hours, like updating an old version of Node.js. It can complete planning steps, modify files, and more in minutes. Today, I'm happy to announce that Jules is entering public beta. Everyone can sign up at jules.google.
As Demis said, we are constantly innovating ways to optimize our models, including improving efficiency and performance. We pioneered the use of diffusion technology to revolutionize image and video generation. Diffusion models generate output by progressively refining noise. Now, we are applying diffusion to text with our latest research model. This helps improve capabilities in areas like text editing, including math and coding, because it doesn't just generate content linearly from left to right; it can quickly iterate on solutions and correct errors during generation.
Gemini Diffusion is a leading text diffusion experimental model that uses parallel generation technology for ultra-low latency. For example, the version of Gemini Diffusion we released today generates content five times faster than our current fastest model, 2.0 Flash-Lite, while having comparable coding ability. Take this math problem, for example. Ready, set, go! The whole process is completed in the blink of an eye. We sped it up before, but this time we'll slow it down a bit.
It's great to see how the model calculates the answer, 39, step by step. This model is currently in small-scale testing. We will continue to explore various methods to reduce latency for all Gemini models and will soon release a faster 2.5 Flash-Lite model. For more information on the future of Gemini, I'll hand it over to Demis.
Thank you, Tulsee. We have been relentlessly pushing the boundaries of Gemini 2.5's thinking capabilities. Our experience building AlphaGo taught us that giving models more time to think improves the quality of their answers. Now, we have further optimized 2.5 Pro by introducing a new mode called Deep Think. It pushes the model's performance to its limits, delivering breakthrough results. Deep Think uses cutting-edge research in thinking and reasoning, including parallel techniques. The performance we've seen so far is stunning.
On the most difficult math benchmark, USAMO 2025, it achieved a very impressive score. It also leads in LiveCodeBench, a high-difficulty benchmark for competitive programming. Gemini has been natively multimodal from the start, so it naturally performs excellently on the main benchmark measuring this capability: MMMU. Since 2.5 Pro Deep Think is at the technological frontier, we need extra time for more advanced safety evaluations and to gather more input from safety experts. Therefore, we will open this feature to trusted testers through the Gemini API to get feedback before a large-scale rollout. Josh will give you more details on our plans for 2.5 Pro Deep Think later today.
Over the past decade, we have done a lot of foundational work for the modern AI era, from our pioneering invention of the Transformer architecture, which is now the basis for all large language models, to building agent systems like AlphaGo and AlphaZero. We will continue to expand the depth and breadth of our foundational research, exploring the next big breakthrough to meet the needs of general artificial intelligence.
Gemini is already an exceptional multimodal foundation model, but we haven't stopped. We are striving to build it into what we call a 'world model'. At that point, it will be able to plan and conceive new experiences by simulating all aspects of the world, much like the human brain. This has always been a personal passion of mine. Early in my career, developing AI for simulation games like Theme Park, I was already exploring virtual worlds.
Over the years, we have been moving in this direction, from training agents to master complex games like Go and StarCraft, to the recently released Genie 2 model, which can generate interactive 3D simulated environments from just a single image prompt. You can already see these capabilities emerging. For example, Gemini can use its understanding and reasoning of the world to represent nature. There's also our advanced video model, Veo, which has a deep understanding of basic physics like gravity, light, and material properties. What's amazing about Veo is its high accuracy and coherence across all frames. Even with a wild prompt, it knows what to do, like this person made of a life raft.
Understanding the physical environment is also crucial for robotics. AI systems need a world model to operate effectively in the real world. We fine-tuned a specialized model, Gemini Robotics, which can teach robots to perform various practical tasks, such as grasping objects, following instructions, and dynamically adapting to new tasks. Everyone here can experience these robots in the AI showcase area.
Building Gemini into a complete world model is crucial for unlocking new types of AI, such as AI that can help with daily life, AI that is intelligent and understands your context, and AI that can plan and take actions across devices for you. Our ultimate vision for the Gemini app is to make it a general-purpose AI assistant, one that is personalized, proactive, and powerful, and a significant milestone on our path to AGI.
This starts with the capabilities we first explored in Project Astra last year, including video understanding, screen sharing, and memory. Over the past year, we have been bringing these capabilities into Gemini Live so more people can experience them today. Of course, we will continue to improve and explore new frontier innovations, like upgrading voice output with native audio to make it more natural. We've improved memory and added computer control capabilities. Let's take a look.
I'm now showing you the latest research prototype of Project Astra. Let's see how it helps me fix my bike. [Project Astra: Action Intelligence + Gemini] Look up the user manual for a Huffy mountain bike online. Found a user manual. What information do you need? Okay, now scroll down to the section on the braking system. Okay, looking for that page. This screw might be stripped. Can you find a tutorial on YouTube to fix it? No problem, opening YouTube.
This video looks good. Put a rubber band over the screw head. Oh, nice, thanks. Check my emails with the bike shop to see what size hex nut I need. Okay, the email from Bicycle Habitat shows you need a 3/8-inch hex nut for your sprocket. I've highlighted the corresponding toolbox for you. I think I need a spare tension screw. Can you contact the nearest hardware store to check if they have it in stock? Okay, contacting them. I'll let you know the inventory details shortly. My brake pads are worn. Please check the user manual again to see if it mentions brake pads. Page 24 of the user manual shows...
Time for lunch? Okay, give me five more minutes. Where were we? You were at page 24, section 2, which explains how to replace the brake pads. Any update on that call? Yes, I just got off the phone with the hardware store. They confirmed the tension screw is in stock. Should I place an order for pickup? Can you show me a few dog baskets suitable for my bike? No problem, happy to help. Take a look at these. I think this one is perfect for Zuco. [Action Intelligence + Gemini: Computer control, highlighting, call assistant, knowledge grounding, native audio, proactive thinking, speaker focus, personalized shopping]
A general-purpose AI assistant will help us handle various tasks. It can handle trivial daily chores and give pleasant new recommendations, thereby improving our efficiency and enriching our lives. We are collecting feedback on these new features from trusted testers and are committed to bringing them to Gemini Live, providing new search experiences, and a Live API for developers, as well as new form factors like Android XR glasses. There will be more detailed introductions later today.
Throughout my career, the core has been about using AI to advance knowledge and accelerate scientific discovery. Google DeepMind has long been dedicated to applying AI to nearly every field of science. In just the past year, we have made huge breakthroughs in multiple areas, including mathematics and life sciences. We built AlphaProof, which can solve Olympiad math problems at a silver medal level. Co-Scientist can collaborate with researchers to help them develop and test new hypotheses. We also just released AlphaEvolve, which can discover new scientific knowledge and accelerate AI training itself.
In life sciences, we built AMIE, a research system that can help clinicians with medical diagnosis. AlphaFold 3 can predict the molecular structures and interactions of all life. And Isomorphic Labs, building on AlphaFold, is dedicated to revolutionizing the drug discovery process with AI, with the potential to help combat various global diseases. In just a few years, AlphaFold has had a profound impact on the scientific community. It has become a standard tool in biology and medical research. Over 2.5 million researchers worldwide use this tool in their critical work.
On the path to AGI, I have always believed that if done safely and responsibly, AI has the potential to accelerate the process of scientific discovery and become the most beneficial technology ever created.
It's incredible to think that just a few years ago, the cutting-edge technology you see today seemed like magic. It's exciting to see this technology powering new experiences in products like Search and Gemini, and collectively helping people in their daily lives. For example, we recently partnered with Aira, a company that helps blind and low-vision people explore the world by connecting them with human 'visual interpreters' via video. Using Astra technology, we designed a prototype to help more people access this kind of assistance. We are continuously collecting user feedback, while Aira's interpreters actively monitor safety and reliability.
With this technology and all our pioneering work, we are building more personalized, proactive, and powerful AI to enrich our lives, accelerate science, and usher in a golden age of discovery and wonder.
I remember as a kid, I would play guitar until it got dark to prove that I could play even if I couldn't see. I was diagnosed with retinitis pigmentosa when I was about four years old. My vision has been deteriorating. But facing a blurring reality, music was my remedy. Playing music made me feel free. If Project Astra could help me live more independently, that would be fantastic.
Take travel, for example. I love to travel because I can go out, see new places, and meet new people. But it's getting harder. Hi. Hi, I'm going to film my surroundings. Can you tell me what's in this green room? The wall you're facing has a sign that says 'Wi-Fi network: TheGreyEagle' and the password is 'livemusic'. Thank you very much. Can you tell me where the microphone stand is? No problem, happy to help. Can I turn on your flashlight? Of course. There are some coiled cables on this wall. The microphone stand should be in the upper right part of the video.
The most powerful thing I can do is get on stage, pick up my guitar, and play. It makes people realize that being able to see or not, being disabled or not, doesn't define everything. If Project Astra can help me, I'm all for it. Have a great show. [Visit goo.gle/AstraVI to join the visual interpreter waiting list]
Thank you, Demis. The potential of Project Astra for accessibility is amazing. This is another exciting example of how AI can advance our mission: to organize the world's information and make it universally accessible and useful.
If there's one product that best embodies our mission, it's Google Search. It's because of Search that we started investing in AI decades ago. And it's through Search that we can make the benefits of AI accessible to everyone, satisfying human curiosity. The Gemini model is making Google Search smarter, more factual, and more personalized. We've made great progress. AI Overviews are a perfect example. Since launching at last year's I/O, the feature has expanded to over 200 countries and regions, serving over 1.5 billion users monthly.
As people use AI Overviews, we find they are more satisfied with search results and search more frequently. In our major markets like the US and India, AI Overviews have driven over 10% growth in related query categories. Most excitingly, this growth momentum is increasing day by day. This is one of the most successful products in Search over the past decade.
AI Overviews are also one of the strongest drivers of visual search growth through Google Lens. Lens searches have grown 65% year-over-year, with over 100 billion visual searches this year. After the release of our latest Gemini model, not only has user query volume increased significantly, but users are also attempting more complex queries. AI Overviews have achieved the search quality and accuracy that users expect, with industry-leading response speed.
To meet the demand for an end-to-end AI search experience, we are launching a new AI Mode. This is a comprehensive reimagining of Search, paired with more advanced reasoning capabilities. You can ask AI Mode longer, more complex queries, like this. In fact, user query length is already 2 to 3 times that of traditional search, and we support multi-turn follow-ups. All of this is launching today as a new tab integrated into Search. I personally use it frequently and find it has completely changed how I search. I'm happy to announce that AI Mode is available to users across the US starting today.
AI Mode will be the first to introduce our cutting-edge technology to Google Search. Starting this week, our most intelligent model series, Gemini 2.5, will be integrated into Google Search, bringing richer capabilities. Now, let me hand it over to Liz.
