A quick note before this article. I put together a guide covering 500+ tested AI image styles and prompts for Midjourney, DALL-E, Stable Diffusion, Leonardo, and more. Every prompt is copy-paste ready and organized by category: cinematic styles, photography looks, fine art, 3D and CGI, fantasy, and dozens more. If photo editing prompts are just the start for you, the full toolkit is right here.

👉 Want 300+ pre-built, tested, copy-paste prompts? Grab the Suno AI Starter Guide here — includes humanizer formulas + stop-slop vocal cues that kill the robotic sound.

Now, back to the article.

Here's the full article rewritten in fresh language:

Something shifted at Google I/O this year.

Past keynotes felt like product announcements. This one felt like a company cashing a decade-long check, showing up with receipts, and daring its competitors to match the tab.

Sundar Pichai walked onto the Shoreline Amphitheatre stage in Mountain View on May 19, 2026, and spent two hours making the case that Google is no longer building AI features into its products. It is building its products around AI, period.

Pichai started with a fact that set the tone for everything that followed. Google committed to being an AI-first company ten years ago. This keynote was about proving that decision was right.

The Gemini app has grown to 900 million people using it every month, up from roughly 400 million a year prior. Google is now handling more than 19 billion AI tokens every single minute across all of its services. And Alphabet has set aside somewhere between $175 billion and $185 billion for capital expenditure in 2026 alone, nearly twice what it spent in all of 2025, with AI infrastructure absorbing the largest share of that.

These figures are not thrown in as background. They are the scaffolding for everything announced after. When a company is spending that kind of money, the features it ships are not experiments. They are commitments.

Here is what those commitments look like in practice.

Before anything else makes sense, you need to understand what Gemini 3.5 Flash is, because it sits underneath every other announcement made today.

Google's new model beats the previous Gemini 3.1 Pro across coding, reasoning, and multimodal tasks while running at four times the output speed of comparable frontier models. That combination of better performance at lower latency is unusual. Usually you trade one for the other. Gemini 3.5 Flash is rolling out today across the Gemini app, Search, the Gemini API, and Antigravity 2.0. Gemini 3.5 Pro, the more powerful sibling, arrives next month.

The practical meaning: every Google product that talks to an AI model just got faster and more capable on the same day, without any action required from users.

Speed upgrades at the model layer matter most when agents are doing multi-step work in the background. A model that is four times faster means a background agent finishes complex tasks in a fraction of the previous time. That connection to agentic workflows is intentional. Google did not build a faster Flash model and then build agents separately. They built the faster model specifically to make the agents viable.

Here is where things start to get real.

Ask anyone who has tried to use an AI assistant for a complex project and they will tell you the same thing. The moment you close your browser tab or step away from the conversation, progress stops. The AI has no memory of what it was doing, no way to continue on its own, and no mechanism to pick up where it left off.

Spark is built to solve that exactly.

Google describes Spark as a personal agent that actively carries out tasks under your direction, running continuously in the cloud on private virtual machines so it keeps going whether your laptop is open or not. You assign it a goal, give it access to the tools it needs, and step away. It does the work.

The live demo on stage showed Spark handling a block party from scratch. A presenter gave it the rough idea, a guest list, and a few constraints. Spark built a spreadsheet to track who was coming, created a shared Google Doc for logistics, sent invitations, monitored RSVPs as they came in, and compiled a list of what each guest offered to bring. None of that required a second prompt. It just ran.

Spark connects to Gmail, Google Docs, and Workspace apps from launch, with MCP-based connections to third-party tools arriving over the summer. The initial rollout goes to Google AI Ultra subscribers in the United States starting next week.

Two additional pieces matter here. First, Spark is also heading to Chrome later this year, so it will eventually be able to take actions in the browser itself, not just within Google's app ecosystem. Second, a new Android feature called Halo shows a live status bar at the top of your phone screen so you can see exactly what Spark is doing at any moment without switching apps.

An agent working while you sleep, visible on your lock screen, connected to your inbox. That is a different kind of software.

Google DeepMind CEO Demis Hassabis took the stage to introduce Omni, and the framing he used is worth paying attention to.

Omni is not a video generator in the traditional sense. It is built on top of what Google calls world models, a class of AI architecture designed to understand physical relationships between objects, not just produce visually convincing output. Most video generation tools today can fool your eyes but not your brain. Things float when they should fall. Water moves like a texture rather than a fluid. Physics breaks in subtle ways that feel wrong even if you cannot immediately name why.

Omni is designed to fix this at the model level. It simulates physical dynamics rather than predicting plausible pixel patterns, which means the resulting videos behave more like footage of real things rather than high-quality hallucinations.

On top of accurate generation, Omni supports fully conversational editing. You watch a video, tell it what you want changed in plain speech, and it applies the edit while preserving everything else. Swap the location, change the outfit, cut to a different angle, slow a specific moment down. No timeline scrubbing, no keyframes, no export-reimport cycle.

You can also generate an avatar version of yourself with your own voice and likeness, and all Omni-generated content is automatically watermarked with Google's SynthID system so it can be identified as AI-created.

The first Omni model, called Omni Flash, is live today for paid Google AI subscribers. YouTube Shorts users get access for free starting this week. Omni Pro follows later. That free YouTube distribution is the strategic move: it puts Google's most advanced video AI in front of an audience far larger than any paid subscription tier could reach.

Pichai called this the biggest Search upgrade in nearly thirty years. That is a strong claim. The substance behind it is real enough to take seriously.

The search box itself has been redesigned. It now expands as you type, and it tries to anticipate what you are building toward rather than just autocompleting the phrase you started. Longer, more conversational queries are treated as natural inputs rather than edge cases. The box also accepts images, files, videos, and Chrome tabs directly alongside text, so you can bring context into a search rather than trying to describe it in words.

Information agents are the bigger structural change. These are background processes that monitor the web continuously for updates related to questions you have already asked. If you ask Search a question about a stock, a product, a news story, or a sports team, an information agent can keep watching that topic and notify you when something changes. It turns Search from a one-time transaction into an ongoing information service. This feature arrives for Pro and Ultra subscribers in the summer.

Search can also now build custom dashboards and mini-apps for recurring tasks. The demo used a weekend family activity planner as an example, where Search generated a dynamic planner populated with local events, weather, and venue information, all organized and updated automatically. These mini-apps are coming later this year for paid tiers in the US.

AI Mode and AI Overviews are merging into a single unified experience. The hard boundary between traditional results and AI-generated answers is dissolving. You start with a regular search, move into a conversation, ask follow-up questions, and circle back to sources, all within the same session without losing the thread. This unified experience is rolling out globally across desktop and mobile today.

The coding platform demo was the moment the room woke up.

Google's Antigravity team gave the platform a live test on stage: build an operating system from scratch. Antigravity 2.0 spun up a fleet of agents, divided the task across them, wrote the code, assembled the OS, and had it running in minutes. The cost for the entire operation came in under $1,000 in tokens. Then, to make the point impossible to miss, they ran a Doom clone on the OS Antigravity had just built.

The new version is 12 times faster than the previous one, which meaningfully changes the economics of running large agentic projects. Token consumption drops, pipeline speed increases, and the total cost of a complex automated workflow becomes significantly more manageable.

Antigravity 2.0 is available globally starting today. The release includes a standalone desktop application, a command-line interface for developers who prefer terminal-based workflows, an updated SDK, and native voice support. All of it runs on Gemini 3.5 Flash.

Google also added Cloud-standard privacy protections to the platform, which matters for enterprise teams who have been hesitant to run sensitive code through a third-party agentic environment.

The Doom demo was clearly aimed at the developers who have migrated toward Claude Code over the past year. Engadget's reporter on the ground said as much in plain language from the keynote floor. Google knows it has ground to recover with developers, and Antigravity 2.0 is the direct answer to that problem.

The most immediately useful announcement for anyone with a Google AI subscription is the pricing change.

A new $100 per month AI Ultra tier is launching alongside the existing structure. It includes five times the usage limits for the Gemini app and Antigravity compared to the $20 Pro plan, 20 terabytes of cloud storage, YouTube Premium, and first access to Gemini Spark starting next week. The existing $250 top-tier plan has been reduced to $200 per month and now offers 20 times the usage limits of the Pro plan.

The pricing math is worth a moment. ChatGPT Pro sits at $200. Claude Max sits at $100. Google's new entry point for its top-tier AI features is now also $100, and the full-featured plan matches ChatGPT Pro's price exactly. That alignment did not happen by accident.

The billing model is also changing. Gemini is moving away from hard daily prompt limits toward a compute-based system where the cost of each interaction is calculated by how complex the task is, what features it uses, and how long the session runs. Lighter tasks use fewer resources. Heavier tasks use more. Users who have been hitting daily limits mid-project will have a more granular and predictable experience.

The Gemini app itself looks different. Google has replaced the previous design with what it calls Neural Expressive, a design language built around fluid animations, richer colors, haptic feedback, and updated typography. Gemini Live has been repositioned in the interface and support for regional dialects is coming in the next few months, including, as demonstrated on stage, a Liverpudlian English option.

Daily Brief is the new agent that earns its keep every morning. It reads your Gmail, scans your Calendar, reviews your task list, and produces a prioritized summary of what the day actually requires from you. It surfaces deadlines buried in email threads, flags meetings you have not prepared for, and suggests actions rather than just listing events. It is rolling out today for all paid subscribers in the US.

The Gemini Mac app is now fully functional for Mac users. The demo showed someone selecting a folder of images and documents in Finder, pressing the Function key, speaking a voice instruction, and having Gemini compose a formatted email using information pulled from those files and sent through Gmail in Chrome. That entire workflow previously required opening multiple applications and copying information between them manually.

Docs Live is the voice-to-document feature. You talk through an idea in natural speech, including pauses, corrections, and repeated thoughts, and Gemini builds a properly structured document from your verbal stream. Versions of this capability are also coming to Gmail and Google Keep.

The app can now also respond with content formats other than text: images, interactive timelines, and narrated video, depending on what the question calls for.

Antigravity 2.0 is globally available right now. The desktop app, CLI, and SDK are all downloadable today. If you have been waiting to try Google's agentic coding platform seriously, the waiting period is over.

Google Flow received the biggest creative tooling upgrade of the keynote. Omni is now integrated into Flow, meaning video generation and editing are directly available inside Google's creative studio. Multi-agent execution is also live, meaning you can run a large number of parallel tasks simultaneously rather than queuing them sequentially. A demo showed 16 different camera angle variants generated from a single source shot, all computed at the same time by 16 separate agents.

Google Flow Music is a new companion tool that allows you to upload audio recordings and use Gemini 3.5 to generate additional musical elements around them. It is an unusual product that sits somewhere between a producer's sketchpad and an AI collaboration tool for audio.

The Gemini API has been enhanced and Google AI Studio now supports native Android development. If you are building Android apps that call Gemini, the integration path is shorter and more direct than it was yesterday.

Code Mender is a new security tool that scans codebases for vulnerabilities and patches them automatically. It is currently in expert testing. The underlying architecture is the same multi-agent platform that powers Antigravity broadly, applied specifically to security instead of general development.

Spark's MCP support for third-party tools arrives over the summer, meaning if you have already built an MCP server for an internal product, Spark will be able to delegate tasks to it without additional integration work on your end.

The hardware reveal came near the end of the keynote and gave the crowd something to talk about beyond software.

Google confirmed its first Android XR smart glasses will ship this fall. The initial lineup comes in two forms: audio-only glasses with cameras, microphones, and speakers built for all-day wear, and a separate variant with an in-lens display that delivers contextual information visible only to the wearer. Samsung and Qualcomm are building the hardware. Gentle Monster and Warby Parker are designing the frames. XREAL is also a platform partner.

Both versions work with Android phones and iPhones. That cross-platform compatibility is a deliberate call. Meta's Ray-Ban AI glasses gained wide adoption partly because they did not exclude iPhone users, and Google is applying the same logic.

The live demo was convincing. A presenter wearing the Gentle Monster frames used voice commands to order coffee through DoorDash, got walking directions delivered through an earpiece, photographed the keynote crowd, and asked Gemini to add generated elements to the shot. The edited image then appeared as a preview on a connected Google Watch.

Maps integration is built in. The glasses can handle navigation, calls, music, and automated tasks, all through voice. No pricing or exact ship date was announced beyond "fall 2026," which means Meta's Ray-Ban lineup, available now starting at $299, holds a practical advantage for the next several months. But the form factor is real, the hardware partners are credible, and the platform story is stronger than anything Google has shown in this space before.

Several other announcements deserve their own attention rather than getting buried in a list.

Universal Cart is Google's AI-powered shopping layer, and it is more sophisticated than the name suggests. It lets you add items from Search, Gemini, YouTube, and Gmail into a single unified cart. But the interesting part is what happens after you add something. The cart proactively tracks price history, watches for restocks, checks for available discounts, and scans for compatibility problems. If you are building a PC and add components that will not work together, the cart catches the conflict and suggests alternatives. Checkout happens through Google Pay directly or redirects to the seller's site. US users get access this summer.

Ask YouTube changes how you find video content. Instead of entering a search term and scrolling through thumbnails, you describe what you want to learn or watch in conversational language and Ask YouTube identifies the best matching content, including specific timestamps for relevant moments within longer videos.

Google Pics is a new image creation and editing tool inside Google Workspace. It generates posters, flyers, and infographics from prompts, and it also allows editing of existing images including removing or resizing elements in both the foreground and background. Every image it generates carries a SynthID watermark.

Stitch is a web design tool that converts rough ideas or voice descriptions into functional interface designs and websites. The target user is anyone who wants to go from concept to something real on screen without writing HTML or using traditional design software.

SynthID, Google's AI content verification system, is expanding significantly. Nvidia, OpenAI, Kakao, and ElevenLabs have all adopted it as a shared standard. It is also rolling out to Google Search through Circle to Search and to Chrome through a right-click verification option on images. The OpenAI adoption is the detail worth holding onto. Two companies in direct competition agreeing on an AI content standard in public suggests both see mandatory regulatory requirements coming and would rather set the standard themselves than have one imposed from outside.

Gemini for Science is a research acceleration toolkit that brings together AI tools for scientific modeling and simulation. The WeatherNext collaboration was cited as a demonstration of the technology's potential, where AI-assisted modeling helped the National Hurricane Center make more accurate predictions about storm track and landfall.

The message delivered from the Shoreline Amphitheatre today was not subtle.

Google has 13 products with over a billion users each, and 5 products with over 3 billion each. Gemini does not need to find new users. It needs to become the default agent layer for people who are already inside the Google ecosystem every day, already trusting Google with their email, their calendar, their search history, and their files.

Spark's ability to run on those private cloud servers while accessing Gmail and Workspace data means Google is building an agent with more context about a user's actual life than any competitor starting from scratch could realistically match. The data advantage is structural, not temporary.

Demis Hassabis said from the stage that AGI is now "just a few years away" and that the right kind of AGI "could promote human progress and flourishing beyond our wildest imagination." That kind of statement is easy to dismiss as keynote theater. The $175 billion capital expenditure sitting behind it is not easy to dismiss.

The competitive picture is straightforward: OpenAI is spending aggressively, Anthropic is moving fast on enterprise, Meta has hardware distribution, and Microsoft has Copilot embedded across Office. Google's answer is that it already has the platform, the users, the data, and the infrastructure. Every announcement today was designed to activate advantages that were already in place rather than build new ones from scratch.

This detail will get four sentences in most roundups. It deserves more.

OpenAI adopting SynthID is the first time Google and OpenAI have publicly shared a technical standard for anything. These two companies are competing for the same users, the same developer relationships, and the same enterprise contracts in almost every category that matters. And yet on AI content verification, they are now aligned.

The most plausible explanation is regulatory anticipation. Both companies see mandatory AI disclosure requirements coming in the EU, the UK, and increasingly in the US. Getting ahead of that with a shared standard they helped build is preferable to having one imposed that neither of them designed. It also signals that neither company sees content verification as a competitive differentiator worth protecting. They both benefit from a world where AI content is identifiable. They just disagree on everything else.

For anyone generating AI-assisted content today, the practical outcome is that the SynthID footprint is about to get much larger. If you are producing content with Google or OpenAI tools for clients or platforms with disclosure requirements, watermarking is effectively already here.

You are reading Roo's Newsletter, so let me be direct about what came out of today for solo creators and independent publishers.

Spark is the most significant shift. If you can hand it a topic, a list of sources to monitor, a document template, and a schedule, and it can research, draft, and queue content for your review while running in the background on Google's servers, the limiting factor in content production stops being time and becomes editorial judgment. That is a meaningful change for anyone operating without a full team.

Daily Brief removes decision fatigue from the first hour of the workday. For a creator managing multiple platforms, publication schedules, and client communication, a morning digest that surfaces what actually needs attention right now is a practical tool rather than a novelty.

Gemini Omni coming to YouTube Shorts for free means video content creation has a lower floor than it had yesterday. If you have been putting off experimenting with video because the tools were too expensive or too technical, that barrier just got meaningfully shorter.

The SynthID expansion is worth tracking carefully. Any content you produce with Gemini is watermarked and identifiable. That is useful transparency for your audience. It also means that platforms and clients developing AI disclosure policies will be able to verify AI involvement in content automatically, which changes the risk calculus around disclosure for creators who have been less than explicit about their AI use.

The Search upgrade matters for distribution. Information agents running continuously in the background, feeding users updates on topics they have already expressed interest in, will reward content that is current, specific, and structured for retrieval rather than optimized for a static keyword position. The SEO playbook does not disappear, but it needs to be read differently.

Google did not ship one big thing at I/O 2026. It shipped twenty interconnected things simultaneously, powered by one faster model, at lower prices than yesterday, with an agent running in the background tying all of them together.

Roo's real take: Gemini Spark is what matters most here, not because it is the flashiest announcement, but because it is the first product in this keynote that changes what you do the night before rather than in the moment you are using it. An AI that is executing tasks while you are asleep is a fundamentally different relationship with software than anything the mainstream market has actually used before. If it works the way the demo suggests, this is the version of AI people were promised three years ago.

One question nobody asked today: if Spark is running on private Google Cloud servers with access to your Gmail, your Docs, your Calendar, and eventually every third-party app you connect through MCP, what is the actual definition of "private" at that scale, and who exactly is in the room when the agent is doing the work?