Google I/O is Google’s annual developer conference, primarily focused on developing applications using Google and open web technologies. The conference content covers Google’s latest technologies, product updates, developer tools, and the company’s future strategic direction. Over the years, Google I/O’s content has reflected the trends in the technology industry, gradually expanding from early focuses on Android and open web technologies to diverse areas such as artificial intelligence, cloud, and hardware ecosystems.
The core philosophy of Google I/O 2025 is “The Gemini Era,” emphasizing the rapid market launch of the most cutting-edge AI models and products, with the goal of making AI accessible to everyone in the world. The main focus of this conference is the deep integration of AI, especially the Gemini model, into Google’s products and services, and promoting its widespread adoption.
Below are the six key highlights regarding Gemini at Google I/O 2025, along with other important announcements:
1. Gemini Model Family Undergoes Comprehensive Evolution
The Gemini model family is evolving at a rapid pace, with materials published quickly from 2024 to 2025, all to bring more new AI experiences to developers.
Sundar Pichai, CEO of Google and Alphabet, mentioned that since the launch of the first generation Gemini Pro model, its Elo rating, which measures model progress, has increased by over 300 points.
- Gemini 2.5 Pro: Dominates various categories on the LMArena leaderboard, performing better on key benchmarks such as reasoning, multimodal processing, code generation, and long context understanding. It includes an experimental Deep Think mode and integrates the LearnLM model co-developed with education experts. New features include Thought Summaries and Thinking Budgets, and it supports Native Audio Output in up to 24 languages.
- Gemini 2.5 Flash: Widely popular due to its fast computing speed and affordable cost, its efficiency has increased by 22%, ranking just behind 2.5 Pro on the LMArena leaderboard. It is expected to be fully available in early June.
- New Gemini Diffusion: Applies Diffusion technology to text editing, including mathematical and code contexts, achieving extremely low latency and generating speeds five times faster than Google’s fastest model, 2.0 Flash-Lite.
- Security: Google emphasizes that Gemini 2.5 is currently the most secure model, with enhanced protection against security threats such as indirect prompt injection.
2. Deep Integration of Gemini into Google’s Core Products
Smart technology and AI are becoming widespread at an unprecedented speed. The total number of tokens processed monthly by Google’s products and APIs has increased from 9.7 trillion last year to over 480 trillion, a 50-fold growth.
- Gemini Applications: Monthly active users have exceeded 400 million, and usage by 2.5 Pro users has increased by 45%. The Gemini App is gradually integrating Project Astra’s capabilities, aiming to become the most personalized, proactive, and powerful universal AI assistant.
- Personal Context: With user consent, the Gemini model can utilize your personal context information from various Google applications (such as Gmail, Google Drive, Google Calendar, etc.) to provide more useful and personalized assistance while ensuring privacy and full control. The new personalized smart reply feature can search your past emails and Drive files, draft specific replies, and learn your common vocabulary, tone, and style. This feature will be available to subscribers later this year.
- Built into Chrome: Gemini will also be built into Chrome for easy use while browsing the web (US users will get early access).
3. Launch and Application of AI Agent Mode
Google views AI Agent as a system that combines the intelligence of advanced AI models with the ability to access tools, capable of performing various tasks under your control.
- Agent Mode: Progress has been made in multitasking capabilities, and it learns task planning through a “teach and repeat” mechanism.
- Application Examples: Agent Mode can help you accomplish more things more efficiently, such as searching, filtering, accessing detailed information on real estate websites, and even booking viewing times. In the AI Mode shopping experience, Agent Mode can even track prices and, under user instructions, complete checkout on their behalf.
- Developer Support: Google is providing Project Mariner’s computer operation capabilities to the developer community through the Gemini API.
- User Oversight: Google emphasizes that Agent Mode is “acting on behalf of” the user under user oversight, not “autonomous decision-making,” and requires user confirmation of key steps.
4. Major AI-Powered Transformation of Google Search
The Gemini model continues to make Google Search smarter, more agentic, and more personalized.
- AI Overviews: Since its launch in 2024, users have expanded to over 1.5 billion, spanning 200 countries and regions.
- New AI Mode: A complete reimagining of search, it integrates all the best AI features and capabilities. With more advanced reasoning capabilities, you can ask AI Mode longer, more complex questions. AI Mode is launching first in the US today. The Gemini 2.5 model version will be introduced in Google Search in the US starting soon.
- Search Live: Leveraging Project Astra’s technology, AI Mode adds the Search Live feature, utilizing both phone camera and screen sharing to obtain useful information.
- Online Shopping: AI Mode also applies to online shopping, adding virtual try-on features, and allowing for price tracking and agentic checkout.
5. Progress and New Tools for Generative Media Models
Previously, Microfusion Technology also wrote an article mentioning that Vertex AI supports multiple media formats, and there have been significant updated versions in less than a month.
- Veo 3: Google’s most advanced video model with native audio generation capabilities. Veo 3 can generate background sound effects and voiceovers.
- Imagen 4: The latest and most powerful image generation model. Imagen 4 focuses more on image details, and better handles text and typography.
- Integration: Both models have been integrated into the Gemini App.
- Flow: Google launched the AI filmmaking tool Flow, based on Veo model technology, developed in collaboration with Hollywood directors. Flow integrates Veo, Imagen, and Gemini, bringing more possibilities for video creators.
- Lyria 2: The text-to-music tool Lyria 2 has also been launched.
- Content Security: SynthID watermarking technology has been extended to images, audio, text, and video, and new SynthID detectors make it easier to identify whether content contains watermarks.
For more detailed image demonstrations, please refer to this article: New Generative AI Media Models Launched: Vertex AI Introduces Imagen 4, Veo 3, and Lyria 2.
6. Android XR Platform and Smart Glasses/Head-mounted Device Layout
- Android XR: Android XR for head-mounted devices and smart glasses, with Gemini as its core product. This is the first Android platform built in the Gemini era.
- Partnerships: The first Android XR head-mounted device, Project Moohan, developed in collaboration with Samsung, is expected to be released later this year. Google will also expand its cooperation with Samsung into the smart glasses field, and partner with eyewear brands Gentle Monster, Warby Parker, and Xreal.
- Application Demonstration: Smart glasses with built-in Android XR were demonstrated on-site, with their cameras and microphones empowering Gemini to see and hear the world. They can perform real-time language translation, obtain information, and even serve as a personal teleprompter. This demonstrates that Gemini’s application on wearable platforms is a strategic extension of Google’s entry into the AR market.
7. New AI Subscription Plans
The conference also announced new subscription plans: Google AI Pro and the new Google AI Ultra.
- Google AI Pro: Monthly fee of $19.99, including higher usage quotas and special features, as well as the Gemini App Pro version previously known as Gemini Advanced.
- Google AI Ultra: Monthly fee of up to $249.99, suitable for users who want to experience Google’s cutting-edge AI, offering the highest usage quotas and earliest access to Google’s various new features and products. The Ultra plan includes Gemini App’s 2.5 Pro Deep Think mode (when ready), Flow and Veo 3 (available today), YouTube Premium, and a large amount of storage space (30TB). New users enjoy a 50% discount for the first 3 months. Google describes the Ultra plan as a VIP pass into the world of Google AI.
8. Infrastructure
- TPU Ironwood: The seventh generation TPU is the first designed specifically to drive thinking and inference-based AI models at scale. Ironwood’s performance is 10 times better than the previous generation, with a total computing power of 42.5 Exaflops per Pod.
Google I/O 2025 clearly announced the full launch of the AI era, with Gemini at
