Google I/O 2025 – The AI Era in Full

Google I/O 2025, the annual developer conference held on May 20‑21, unequivocally confirmed the dominant role of artificial intelligence in the company's strategy. This year's event was saturated with AI announcements, underscoring Google's transformation into an organization for which AI is a priority. Sundar Pichai, Google's CEO, described the current phase as a “new phase of AI platform change”, where decades of research become reality for people worldwide. The conference sent a clear signal that artificial intelligence is no longer merely an add‑on, but the foundation of the company's future, with a vision “AI Everywhere” – from search to developer tools to new forms of interaction.

Google aims for AI to become an omnipresent “collaborator and superpower”, automating the “endless grind” and integrating into every aspect of digital life, even extending beyond screens into the physical world. This reflects a fundamental strategic shift. For years Google championed a “Mobile‑First” strategy, which was clearly visible at previous I/O conferences. However, this year’s I/O 2025, with almost every announcement focused on AI and even spinning off the Android Show as a separate event, clearly signals a fundamental paradigm shift. It means that AI is no longer just a feature in mobile apps, but becomes the foundational layer upon which all future Google products and experiences are built, regardless of device. This shift aims to maintain a leadership position amid growing competition and define a new era of digital interactions.

It is worth noting that Google employs a dual‑track strategy in AI delivery. On one hand the company announces free access to Gemini Live and Gemini 2.5 Flash for everyone, as well as free upgrades for students. On the other hand, it introduces paid subscriptions Google AI Pro and Ultra for more advanced features and higher limits. This distribution model suggests that Google aims for broad adoption of its AI technologies and building an extensive ecosystem by offering basic yet powerful tools for free. At the same time the company monetizes advanced use cases and higher limits, targeting professionals and enterprises that need greater power and specialized features. This approach also serves as a response to OpenAI’s business model, where access to GPT‑4o is largely paid.

---

New AI Models and Tools: The Heart of Innovation

The Google I/O 2025 conference delivered a series of groundbreaking announcements regarding AI models and developer tools that form the foundation of a new “AI‑first” era.

Gemini 2.5: Greater Power and Availability

Google showcased significant enhancements to its Gemini model family. It introduced Gemini 1.5 Flash, a model optimized for speed and high‑frequency tasks, available via the Gemini API in Google AI Studio. An upgraded version, Gemini 1.5 Pro, offers an impressive context window of 1 million tokens in the public beta, with the possibility to expand to 2 million tokens for developers on the waitlist.

Both models, Gemini 2.5 Pro and Flash, gained new capabilities, including advanced safety features. Google claims that Gemini 2.5 Pro is currently the leading model on the WebDev Arena and LMArena ranking lists. The model has been enriched with LearnLM, making it a leading learning model, surpassing competitors in every learning‑rule category. Deep Think, an experimental enhanced reasoning mode for very complex mathematical and coding tasks, was also announced for 2.5 Pro. The new preview version of Gemini 2.5 Flash offers better performance on coding and complex reasoning tasks, optimized for speed and efficiency. The Thinking Budgets feature, introduced in 2.5 Flash and extended to 2.5 Pro, allows developers to control costs by balancing latency and quality, as well as the number of tokens the model uses for “thinking”.

Gemini 2.5 Flash is already available to everyone in the Gemini app. Google also introduced new paid subscription plans: Google AI Pro ($19.99 USD/month) offering higher limits and access to tools such as Flow and NotebookLM, and Google AI Ultra ($249.99 USD/month) with the highest limits, access to the most advanced models and premium features, as well as 30 TB of storage and YouTube Premium. Additionally, students in the US, Brazil, Indonesia, Japan, and the United Kingdom qualify for a free Gemini upgrade for an academic year. The introduction of multiple Gemini variants (Flash, Pro, Ultra) and differentiated subscription plans (free, Pro, Ultra, student) indicates a deliberate market‑segmentation strategy. Google does not offer a single universal solution, but tailors its AI models and services to various user needs – from fast, lightweight tasks (Flash) to complex, deep reasoning (Pro with a large context window and Deep Think) and premium features (Ultra). This strategy enables the company to reach a broad user base, from beginners to advanced developers and enterprises, maximizing adoption and potential revenue.

Gemini Models and Their Use Cases:

Here is an overview of Gemini models, along with their key features and availability:

• Gemini 1.5 Flash:
  • Main Features: Lightweight, fast, for high‑frequency tasks.
  • Context Window: Optimized for speed (no specific token window as in other models).
  • Availability: Public preview in over 200 countries.
• Gemini 1.5 Pro:
  • Main Features: Enhanced, offers advanced reasoning, supported by LearnLM (makes it a leading learning model).
  • Context Window: 1 million tokens (public preview), 2 million tokens (waitlist for developers).
  • Availability: Public preview in over 200 countries.
• Gemini 2.5 Pro:
  • Main Features: Leader on WebDev Arena and LMArena rankings, announced experimental reasoning mode Deep Think for complex tasks.
  • Context Window: 1 million tokens (future expansion to 2 million).
  • Availability: Generally available in Google AI Studio and Vertex AI soon.
• Gemini 2.5 Flash:
  • Main Features: Faster, better performance on coding and complex reasoning, optimized for speed. Introduced Thinking Budgets, allowing cost and quality control.
  • Context Window: Optimized for speed.
  • Availability: Available to everyone in the Gemini app.
• Gemini Nano:
  • Main Features: Google’s most efficient on‑device model, providing low latency and increased privacy.
  • Context Window: Runs on device.
  • Availability: Available on Pixel 8 Pro, Samsung Galaxy S24 Series, and in the Chrome (desktop) browser.
• Gemini Diffusion:
  • Main Features: Research model that generates text and code from “random noise” (similar to image/video generation models). It is 10‑15 times faster at coding than autoregressive models.
  • Context Window: (Diffusion technique, not applicable to a typical context window).
  • Availability: Research model, under development.

Gemma Model Family (Open Source):

• Gemma (general):
  • Main Features: Open model family, including CodeGemma (for code), RecurrentGemma, and PaliGemma (for visual‑language tasks).
  • Gemma 2 (27 B parameters): Designed to outperform models twice its size.
  • Gemma 3n: Fast and efficient multimodal model, optimized for mobile devices.
  • Availability: Available, including in Google AI Studio and Google Cloud.

Multimodal Generation: Image, Video, Audio

Google continues investing in generative AI, introducing new models capable of creating images, video, and audio:

• The latest Imagen 4 model offers extraordinary detail clarity, excels at photorealistic and abstract styles, and significantly improves text rendering and typography. It is available in the Gemini app and Vertex AI, and its fast version (up to 10× faster than Imagen 3) will be available soon.
• Veo 3, the new generation of Google’s AI video generator, stands out with the ability to generate video with native audio, including dialogues.
• Flow is a new AI video‑creation tool that leverages Veo, Imagen, and Gemini to generate and stitch AI‑generated video clips, offering control over camera and scenes. Flow is available to Google AI Pro and Ultra subscribers.
• Access to the Music AI Sandbox, powered by Lyria 2, has also been expanded, offering powerful composition and vocal generation capabilities, as well as an interactive Lyria RealTime model for real‑time music creation.
• A new research model, Gemini Diffusion, generates text and code by transforming random noise, similar to image and video generation models. It is considerably faster (10‑15× faster than autoregressive models) and particularly strong at coding.

Google clearly aims for all its generative models (text, image, video, audio) to be multimodal and integrated with each other. The fact that Veo 3 generates audio and Imagen 4 improves text rendering shows that the boundaries between modalities are blurring. Gemini Diffusion, using diffusion for text and code, further underscores this convergence, suggesting that shared architectures can be applied to different data types. This means future AI applications will seamlessly transition between modalities, offering more natural and versatile user experiences.

Open Ecosystem and Tools for Developers

Google is actively expanding its open AI ecosystem, releasing new models and tools for developers. The open model family Gemma now includes CodeGemma and RecurrentGemma, as well as PaliGemma for visual‑language tasks. Gemma 2 (27 B parameters) was also showcased, designed to outperform models twice its size. Gemma 3n is another fast and efficient multimodal model, optimized for mobile devices.

The Gemini API received enhancements, including support for parallel function calls and video frame extraction. Context Caching was introduced for large prompts, and the Model Context Protocol (MCP) for easier integration with open‑source tools.

Among the new developer tools are:

• Google AI Studio: With an improved UI, integrated documentation, new apps, and a “Generate Media” tab.
• Agentic Colab: Colab will become a fully agentic experience, allowing users to describe goals, and the AI will operate in notebooks, fixing bugs and refactoring code.
• Stitch (UI design): New AI tool for generating high‑quality UI designs and corresponding frontend code from natural‑language descriptions or images.
• Jules (coding agent): Asynchronous AI agent for GitHub repositories that can fix bugs, add documentation, and build new features, running in the background. It is already in public beta and Google emphasizes it does not train on private code.
• Gemini in Android Studio: Evolution of Studio Bot, now stable and available in over 200 countries, with future support for multimodal inputs.
• Gemini Nano & AICore: Gemini Nano, Google’s most efficient on‑device model for tasks, runs directly on mobile devices, providing low latency and increased privacy. AICore manages on‑device models.
• Checks: AI‑based compliance platform simplifying privacy and compliance workflows for apps.

Repeated mentions of Gemini Nano in Chrome and on mobile devices (Pixel 8 Pro, Galaxy S24 Series), as well as Gemma 3n optimized for mobile devices, indicate Google’s strategic emphasis on on‑device AI. Running AI directly on the device has key implications for privacy (data never leaves the device), speed (low latency), and availability (independence from internet connectivity). It also responds to the growing demand for personalization and security in the AI era.

Announcements such as Agentic Colab, Stitch, and Jules signal a fundamental shift in how developers will build software. Instead of writing every line of code, developers will increasingly “orchestrate” AI agents that automatically generate user interfaces, fix bugs, write tests, and manage repositories. This shift from “coding” to “managing AI to generate code” has the potential to dramatically boost productivity, but also demands new skills and workflows from developers.

---

AI in Google Products: Revolution in Everyday Life

Artificial intelligence is becoming an integral part of Google products, changing how users interact with technology in everyday life.

The Future of Search: AI Mode and AI Overviews

AI Overviews, generative summaries in Google Search, scaled to 1.5 billion monthly users in 200 countries. Google is rolling out “AI Mode” as a generally available option in the US, turning traditional search into a conversational interaction with an expert.

New capabilities of AI Mode include:

• Deep Search: Provides deeper answers to complex questions, leveraging advanced research capabilities.
• Search Live (integration with Project Astra): Allows users to converse with the search engine in real time about what they see through the camera.
• Agentic Capabilities (Project Mariner): Enables automation of browser tasks such as ticket purchases or restaurant reservations.
• Data analysis and graphics: AI Mode will analyze complex data sets and create custom graphics, initially for sports and financial queries.
• Shopping experiences: The new AI Mode shopping experience combines advanced AI with the Shopping Graph, helping with inspiration, consideration, and finding the right product.
• Virtual try‑on and agentic checkout: Users can virtually try on clothing and use the “buy for me” feature, which automatically places an order when the price drops to the desired level.

The shift to “AI Mode” and the dominance of “AI Overviews” is not merely an evolution of the search engine, but a fundamental redefinition. Instead of providing a list of links, Google aims to deliver direct, comprehensive answers. While Google claims this will increase the number of searches and clicks, studies show click‑through rates have dropped by almost 30%. This raises serious concerns for publishers and content creators whose businesses rely on search traffic. Google, becoming the “answer” rather than the “signpost,” could “steal” content without offering much in return. This strategic move could revolutionize the web, but also challenges its open nature and the business models of many websites.

Key AI Features in Google Products:

Artificial intelligence is becoming an integral part of Google products, changing how users interact with technology in everyday life.

• Google Search:
  • AI Mode: Conversational search that provides direct answers and replaces traditional SERP listings.
  • Deep Search: Provides deeper answers to complex questions, using advanced AI research capabilities.
  • Search Live (integration with Project Astra): Enables real‑time conversation with the search engine based on images from the device’s camera.
  • Agentic Shopping / Buy for Me: Automatically tracks prices and purchases products when they reach the desired price.
  • Virtual Try‑On: Allows users to virtually try on clothing based on their own photo.
• Gemini App:
  • Gemini Live (with camera/screen): Free multimodal voice assistant that can use context from the user’s surroundings (camera image, on‑screen content).
  • Integration with Google Apps: Gemini will connect with Maps, Calendar, Tasks, and Keep to offer personalized actions.
• Gmail:
  • Personalized Smart Replies: Intelligent replies tailored to the user’s context and previous messages/files from Google Drive.
• NotebookLM:
  • Audio/Video Overviews: Generating audio and video summaries from documents and files.
• Google Meet:
  • Speech Translation: Real‑time speech translation that preserves voice quality and expression.
• Android Studio:
  • Gemini in Android Studio: AI‑powered coding assistant that speeds up Android app development.
• Chrome Desktop:
  • Gemini Nano in Chrome: On‑device AI running directly in the Chrome browser, enhancing privacy and speed.

Intelligent Assistants and Productivity

Google unveiled Project Astra, a multimodal AI assistant prototype that can “see” the surroundings through a phone’s camera and perform tasks without explicit commands. Improvements include more natural speech, enhanced memory, and computer control. Prototypes include a conversational tutor and assistance for visually impaired users.

The Gemini Live feature with camera and screen sharing is now free for all Android and iOS users. In the coming weeks, Gemini Live will become more personalized through integration with Google apps such as Maps, Calendar, Tasks, and Keep, enabling actions during the conversation. Gmail will receive personalized “Smart Replies” that consider the user’s context and tone from previous messages and Drive files. NotebookLM now offers flexibility in the length of “Audio Overviews” and will soon introduce “Video Overviews”. Google Vids is now available to Google AI Pro and Ultra users. Sparkify is an experiment that turns queries into short animated videos.

The evolution from “reactive” chatbots to “proactive” agents, such as Project Astra or agentic capabilities in AI Mode, is a key trend. Google not only answers queries but aims for AI to understand user goals, plan, and act on their behalf (e.g., booking tickets, shopping, organizing travel). This means AI will become increasingly embedded in user workflows, anticipating needs and performing complex tasks autonomously, leading to a significant boost in productivity and convenience.

---

AI Beyond the Screen: Android XR and Google Beam

Google clearly envisions the future of AI not only in software but also in physical devices that blur the line between the digital and real worlds. The company is heavily developing the Android XR platform for mixed‑reality devices. The first device, the Samsung Project Moohan headset, is set to deliver immersive experiences on an “infinite screen” later this year. Gemini running on Android XR glasses was demonstrated in real‑world scenarios such as messaging, meetings, and guidance. The glasses will be equipped with microphones, a camera, and speakers, allowing Gemini to obtain context from the user’s surroundings. Live language translation between two people using Android XR glasses was also showcased, highlighting the potential to break language barriers.

Project Starline, a research project on 3D video calls, is evolving into a new platform called Google Beam. Google is partnering with Zoom and HP to bring the first Google Beam devices to market later this year. The goal is to create more natural and intuitive virtual conversations without the need for headsets. Real‑time speech translation in Google Meet is already available, providing real‑time speech translation while preserving voice quality, tone, and expression.

Announcements regarding Android XR and Google Beam clearly show that Google sees the future of AI not only in software but also in physical devices that blur the line between the digital and real worlds. AI in XR glasses that “sees” and “hears” the environment, and 3D communication in Google Beam that eliminates the need for headsets, are steps toward “spatial computing” and immersive experiences. This means AI will increasingly help users interact with the physical world, not just screens, which has huge implications for work, education, and social life.

---

Competition and Challenges: AI Landscape

The AI landscape is intensely competitive, and Google I/O 2025 highlighted both Google’s strengths and the challenges facing the industry.

Brief Comparison with Key Competitors

• Gemini vs. GPT‑4o: Google Gemini 2.5 Pro with 1 million (soon 2 million) token context far exceeds GPT‑4o’s 128 thousand tokens. Google claims Gemini 2.5 Pro is better at reasoning, maintaining context, and solving AI problems. Gemini 2.5 Pro is also free with limits, whereas GPT‑4o requires a paid subscription. GPT‑4o stands out for real‑time multimodality (voice, vision, text), although Gemini 2.0 Flash also shows an advantage in multimodal understanding on some benchmarks.
• Veo 3 vs. Sora: Veo 3 stands out with the ability to generate native audio, including dialogues, which Sora lacks. Veo 2.0 (previous version) supports up to 4K resolution, while Sora generates up to 1080p. Google Flow, an AI video‑creation tool, combines Veo 3 with Imagen 4. OpenAI’s Sora is closed‑source, available through ChatGPT+, whereas Veo is available via Google platforms (Vertex AI, YouTube).
• Jules vs. GitHub Copilot: Jules is an asynchronous coding agent that runs in the background, performing tasks such as fixing bugs, adding features, and writing tests. Unlike Copilot, which mainly provides intelligent autocomplete, Jules lets developers assign tasks and return to them later. Google emphasizes that Jules does not train on private code.
• Flow vs. Adobe Firefly: Flow is an AI video‑creation tool that uses Google models (Imagen, Veo, Gemini). Adobe Firefly, while offering new image and video models (Firefly 4 and 4 Ultra) and the ability to use OpenAI and Google models in Firefly, does not have native audio generation in video as Veo 3 does.

Comparison of Generative Video/Image Models:

Here is a comparison of generative video and image models from Google and competitors:

• Google Imagen 4:
  • Generation Type: Image
  • Key Features: Exceptional detail clarity, photorealism, improved text rendering and typography.
  • Status: Available in the Gemini app and Vertex AI.
• Google Veo 3:
  • Generation Type: Video
  • Key Features: Generates video with native audio (including dialogues), camera control, outpainting, adding/removing objects.
  • Status: Available in the Gemini (Ultra) app, Vertex AI.
• Google Flow:
  • Generation Type: Video (tool)
  • Key Features: Combines Veo, Imagen, Gemini to create and edit AI‑generated videos, control over characters, scenes, styles.
  • Status: Available in the Gemini (Pro/Ultra) app.
• OpenAI Sora:
  • Generation Type: Video
  • Key Features: Generates coherent video clips, integrating latent diffusion with Transformers, automatic metadata tagging.
  • Status: Closed source, available through ChatGPT+.
• Adobe Firefly (models):
  • Generation Type: Image/Video
  • Key Features: New Firefly 4/4 Ultra models for detail and realism; ability to use OpenAI and Google models.
  • Status: Available, including in public beta.

Discussion of Controversies and Challenges

Google I/O 2025 is an AI innovation festival, but it also reveals tensions associated with the rapid pace of development. While AI Overviews increase Google usage, there are concerns about their accuracy and a drop in click‑through rates for publishers (almost a 30% decline over a year). This raises questions about publishers’ business models and “theft” of content by Google without proper compensation.

The rise of AI‑driven personalization raises privacy concerns and a “Big Brother” effect, as AI understands users at a level they were unaware of. Google wants users to allow Gemini models to read data from Google services for personalization. Despite progress, Google acknowledges that generated media can contain “slop”, highlighting the need for tools to detect AI‑generated content, such as the SynthID Detector.

AI development requires massive infrastructure investments (e.g., Ironwood TPU). Running AI applications can be costly, and many organizations still struggle to assess ROI on AI. Growing concerns about AI regulation, including copyright issues and liability for AI agents’ actions, are becoming increasingly visible.

Google I/O 2025 is an AI innovation festival, but it also reveals tensions. On one hand, the company aggressively embeds AI into every product and develops advanced models. On the other hand, serious questions arise about ethics (privacy, bias, AI autonomy), responsibility (accuracy of AI Overviews), and sustainability (the massive energy consumption of AI). Google is trying to address these issues (e.g., SynthID, Jules privacy), but the scale and pace of change suggest that regulations and societal adaptations will need to keep up with technological progress. This shows that “AI‑first” does not mean “problem‑free”.

---

Summary and Outlook

The Google I/O 2025 conference was a clear confirmation that Google places AI at the core of its strategy. AI dominance was evident at every turn – from the development of advanced Gemini models, through their deep integration in consumer products (Search, Gmail), to new platforms (Android XR, Google Beam) and developer tools (Jules, Stitch, Agentic Colab). There is a pronounced focus on multimodality, AI proactivity, and its ability to act on behalf of the user.

What this means for the future of technology and daily life are shifting interactions. AI will increasingly “understand the world”, becoming a proactive “collaborator”, not just a tool. This means less “searching” and more “acting” via AI. The development of Android XR and Google Beam heralds a future where AI extends beyond screens, integrating with the physical reality and enabling more natural, immersive interactions. The developer transformation, evident in tools like Jules and Agentic Colab, points to an evolution of the developer role, who will increasingly manage AI agents rather than write code from scratch.

The rise of advanced AI also brings serious challenges regarding privacy, ethics, content ownership, and impact on traditional business models. Google will need to balance innovation with responsible development. Historically, “Google It” was synonymous with information search. However, the introduction of “AI Mode” and “Agentic Capabilities” suggests Google aims to shift this paradigm to “Google Do It”. Instead of searching for information, users will delegate tasks to AI, which will execute them autonomously (e.g., reservations, purchases). This means Google wants to dominate not only access to information but also the execution of online actions. If this vision materializes, it could strengthen Google’s position as the primary gateway to the internet and digital services, but also raise further concerns about centralization and control over the user experience.

Google I/O 2025 not only showed what Google is building but also signaled the direction the entire tech industry is heading. We are witnessing deep AI integration that will revolutionize how we work, communicate, and interact with the world. It is an exciting yet demanding time, where the focus will be not only on creating powerful models but also on building them responsibly and for the benefit of everyone.