Understanding the 2025 Google I/O Developer Conference

On May 21, 2025, the Google I/O Developer Conference arrived as scheduled. This highly anticipated technology event once again brought many surprises to global audiences.

From brand-new AI models to innovative product applications, from breakthroughs in visual generation technology to the reshaping of the search experience.

At this conference, Google fully demonstrated its latest achievements and strategic layout in the field of artificial intelligence .

1. Google AI Ultra Membership: Strategic Layout

At this conference, Google launched a major initiative - the Google AI Ultra membership service priced at $249.99 per month .

This membership package includes almost all the new products and services released by Google this time , such as Gemini 2.5 Pro Deep Think, Veo3, Project Mariner , etc., and also integrates membership benefits such as NotebookLM and YouTube .

This is not only an important step for Google in product integration, but also a strategic manifestation of its top-level promotion of the popularization and application of AI technology .

Currently, the membership service is half price for the first three months, only US$124.99 per month , which is undoubtedly very attractive to users who pursue cutting-edge technology experience.

2. AI Model: Double Breakthrough in Performance and Functionality

1. Gemini 2.5 Pro: Strong performance of the all-around champion

Gemini 2.5 Pro is undoubtedly one of the star products of this conference .

It has demonstrated excellent performance in multiple dimensions such as language, reasoning, and code , directly crushing all the test items of LMArena . In particular, in the WebDev coding test, Gemini 2.5 Pro achieved a ceiling-level score, which increased by 142 Elo points compared to the previous generation, firmly ranking first.

What's even more amazing is that it also successfully cleared "Pokémon: Blue", completed the collection of eight badges and the challenges of the Four Heavenly Kings and the Champion. This achievement was jokingly called "Artificial Pokémon Intelligence (API)" by the audience.

2. Gemini 2.5 Flash: Efficient performance for speed players

If the Gemini 2.5 Pro is the all-around champion, then the Gemini 2.5 Flash is the representative of speed players .

Google calls it the " most efficient working model " with significant improvements in reasoning, code, and long context. Its overall performance on the LMArena list is second only to Gemini 2.5 Pro.

The new version of Gemini 2.5 Flash is expected to be officially launched in early June . Users can currently experience the preview version in AI Studio, Vertex AI and Gemini App.

3. Gemini 2.5 Pro Deep Think: Super-powerful deep thinking

Google has added a new Deep Think mode to the Gemini 2.5 Pro , which makes it perform amazingly well on extremely difficult math and programming benchmarks.

This mode is similar to OpenAI's o1 pro, which uses a longer reasoning time in exchange for higher quality. Currently, Gemini 2.5 Pro Deep Think is still in the security evaluation stage and is only open to some trusted testers.

4. Gemini Diffusion: Experimental Text Diffusion Model

Google also demonstrated a cutting-edge research - Gemini Diffusion .

Unlike traditional models that generate content by predicting the next token, Gemini Diffusion generates content in parallel by gradually refining the noise , showing extremely low latency .

The live demo version of Gemini Diffusion builds five times faster than Flash Lite 2.0, with nearly equal code performance.

3. Gemini products: multimodal interaction and personalized experience

1. Gemini Live: A new experience of multimodal interaction

Gemini Live is a new product from Google that integrates multimodal interaction functions . It originated from last year's research project Project Astra.

Users can talk to the AI ​​assistant by pointing at the phone camera or anything on the screen . For example, when the tester pointed the camera at the street, Gemini Live not only corrected the user's misperception that a garbage truck was a luxury convertible, but also taught him common sense that street lamps are not tall buildings.

Starting today, Gemini Live’s visual question-and-answer feature will be fully available on Android and iOS platforms .

2. Personal Context: Personalized and thoughtful service

After obtaining user authorization, the Personal Context function can connect search history, Gmail, Drive and other information in Google applications to truly realize personalized services.

For example, it can generate personalized exam questions for users based on their physical exam times, notes, professor materials, and even handwritten notes.

3. DeepResearch and Canvas Creation: Upgrading Deep Research and Creation

The DeepResearch feature now supports users to upload their own files for in-depth research, and Canvas has also been upgraded.

Users can convert research reports into web pages, infographics, and even podcasts in 45 languages ​​with one click. In addition, Canvas also supports front-end preview of code and can share the code with friends.

4. Gemini appears in Chrome: Summary of AI in web plugins

When browsing the web, Gemini can directly understand the content of the current page and answer user questions , which is equivalent to providing users with an AI-summarized web plug-in. This function will greatly improve the efficiency of users' information acquisition during web browsing.

5. Gemini Agent Mode: Let AI do the work for you

For C-end users, Google introduced a new Agent Mode in the Gemini application .

Users can tell Agent Mode complex task requirements, such as finding a house, and Agent Mode will automatically search for qualified houses on real estate websites such as Zillow and call Mariner to screen. When a desired house is found, it can also automatically submit a form to make an appointment to view the house and continuously update the house information.

Currently, Gemini Agent Mode is still in the experimental stage and is expected to be launched in an experimental version to subscribers soon.

4. Visual Generation: From Static to Dynamic

1. Flow: An innovative tool for AI filmmaking

Flow is a new AI film-making tool launched by Google at this conference . It integrates the capabilities of Veo, Imagen and Gemini, and directly competes with Sora .

Users can upload their own pictures, or use the built-in Imagen to generate materials, and then generate a high-quality AI video with one command.

After the video is generated, users can also edit it directly and even add new shots . However, Flow is currently only available to users in the United States , and users can only use Google's own Image to generate it , and cannot upload pictures themselves, which limits its application scenarios to a certain extent.

2. Veo3: A breakthrough in Vincent Video

Veo 3 is another masterpiece of Google in the field of real-life video. It has significantly improved the image quality and realism compared with the previous generation , especially in the deeper understanding of physical laws , such as gravity, lighting, materials and other real-life laws.

The biggest breakthrough of Veo 3 is the addition of native audio generation function for the first time , allowing users to include background sound effects, ambient sounds and even character dialogues in AI-generated videos.

Veo 3 was launched on the Flow platform on the day of its release, but is currently only available to Google AI Ultra members . The cost of generating videos is high, with 150 points required to produce one video, and Ultra members only have 12,500 points per month.

3. Imagen4: A qualitative change in text-generated images

Imagen 4 is Google 's latest generation of text-generating image models , with a qualitative leap in image quality, richer colors and more refined details.

Imagen 4 has also made significant improvements in text embedding , and can accurately generate text content, select appropriate font styles based on the content, and automatically adjust spacing and layout.

Currently, Imagen 4 has been fully launched and ordinary users can use it on the Whisk platform.

5. Google Search: AI reshapes the search experience

1. AI Overviews: Widespread Application of AI Summary

Google Search’s AI Overviews feature has achieved remarkable results over the past year, with more than 1.5 billion users currently using the feature each month.

AI Overviews generates a brief description generated by the Gemini model at the top of the results when users search, with a reference to the source of the information. With the application of the Gemini 2.5 series of models, the results generated by AI Overviews have been further improved in terms of accuracy and coverage.

Currently, this feature has been launched in more than 40 countries and regions including the United States and India, and supports multi-language queries. Since the launch of AI Overviews, users have increased their investment in search, with daily search volume increasing by an additional double-digit percentage.

2. AI Mode: Comprehensive reconstruction of the search paradigm

AI Mode is the end-to-end AI search experience launched by Google at this conference .

With AI Mode enabled, users can ask longer and more complex questions without having to condense them down to keywords as they would in the past.

AI Mode can provide personalized search results based on the user's search history and even Gmail information (with user consent) . In addition, AI Mode also has the following highlights:

- Deep Search : For issues that require in-depth research, AI Mode can issue dozens or even hundreds of search requests simultaneously and aggregate the information into an expert, referenced report.

- Complex data analysis and visualization : AI Mode can generate tables and charts to help users understand data more intuitively. In the future, it will also support more detailed analysis of sports and financial issues.

- AI helps you do things : AI Mode integrates the capabilities of Project Mariner and can help users complete some practical operations, such as purchasing game tickets, booking restaurants, and making appointments for services.

- AI shopping experience : AI Mode can recommend suitable products based on the user's description and provide purchase links and precautions. In addition, it also has a virtual fitting function, where users only need to upload photos to see how the clothes will look on them.

Currently, AI Mode is open to all users in the United States as an experimental new feature. The launch of this feature marks Google's comprehensive transformation from the traditional "search result list" to the "conversational report" search paradigm.

6. Agent system: integration of automation and intelligence

1. Project Mariner: AI-powered automated browser assistant

Project Mariner is an important achievement of Google in Agent technology .

Since its launch as an early prototype last December, Mariner has learned to multitask, able to oversee up to 10 tasks simultaneously.

In addition, it also introduces the " Teach and Repeat " function, which allows Mariner to learn the process and get started directly when encountering similar tasks with just one demonstration.

This feature gives Mariner the shadow of RPA (robotic process automation) , providing broad possibilities for future automation applications.

Currently, Mariner has provided its computer-based capabilities to developers through the Gemini API , and it is expected to open this capability to developers on a larger scale this summer.

2. Jules: Exploration of AI Programming Agent

Jules is an AI programming agent released by Google in advance before this conference and is currently in the global testing phase.

Users can connect to GitHub and automatically pull and submit code through Jules. However, Jules must be connected to GitHub to use , which may not be very friendly to novice developers.

7. Other highlights: Collaborative innovation of hardware and software

1. NotebookLM: A standalone application for AI podcasts and knowledge organization

NotebookLM is one of the hottest projects last year and this year , and it has set off a trend of AI podcasts .

At this conference, Google announced that it will officially launch an independent application and land on Android and iOS platforms . Users can use NotebookLM anytime and anywhere to organize notes and knowledge, whether on mobile phones or on the web.

2. Gemini is integrated into the Android family bucket: seamless connection of multiple devices

Currently, users can quickly start Gemini by pressing the power button .

In the coming months, Gemini will also be available on smart watches, car dashboards and even TVs , enabling seamless connection and interaction across multiple devices.

3. Project Moohan headset: Android XR device jointly developed with Samsung

The Project Moohan headset is the first Android XR device jointly developed by Google and Samsung and is expected to be officially launched later this year.

This headset will provide users with an immersive virtual reality experience , further expanding Google's layout in the XR field.

4. Android XR smart glasses: smart interactive devices worn all day

Google showed off a lightweight, all-day-wearable pair of Android XR smart glasses at the conference , with built-in cameras, microphones, speakers, and even an optional in-lens display.

Users can use the glasses to realize a variety of intelligent interactive functions, such as: identifying surrounding objects, answering questions, playing music, navigation, identifying coffee brands and finding coffee shops , as well as real-time translation functions .

Google has also collaborated with brands such as Gentle Monster and Warby Parker to further expand the market for smart glasses.

5. Google Beam: An upgrade to 3D video calling technology

Google Beam is an upgraded version of the 3D video calling technology previously launched by Google .

It captures the user from different angles through an array of six cameras , then fuses these video streams into a realistic 3D experience through AI, and can track the head in real time on a 3D light field display with millimeter-level accuracy and 60 frames per second. HP is expected to launch the first batch of Google Beam devices later this year.

6. Google Meet real-time voice translation: barrier-free communication

Google Meet's real-time voice translation technology enables people of different languages ​​to have natural and fluent conversations, and the translation can match the speaker's tone, speed and even expression.

Currently, this feature can be used directly in Google Meet and supports English and Spanish. More languages ​​​​and enterprise versions will be launched in the future.

7. TPU Ironwood: The seventh generation of TPU performance leap

TPU Ironwood is the seventh-generation TPU launched by Google , specially designed for large-scale applications of AI thinking and reasoning.

Its performance is 10 times higher than the previous generation and is expected to be launched on Google Cloud later this year. The launch of this technology will provide strong hardware support for the further development of AI applications.

8. SynthID Digital Watermark: Copyright Protection for AI-Generated Content

In order to address the problem of AI-generated content being difficult to identify , Google upgraded its SynthID technology .

The new SynthID Detector can detect whether a SynthID tag is contained in an image, audio, text, or video , even in a small portion. This technology will help protect the copyright of AI-generated content and prevent its misuse.