Google Goes All-In on Agents at I/O 2026: Spark, Search, Omni

Google's Search box hasn't changed in any fundamental way since the late 1990s. You type, it guesses, you click. The underlying mechanics evolved over 25 years, but the paradigm stayed intact. At its annual I/O developer conference on May 19, 2026, Google announced it was replacing that paradigm entirely.

This wasn't a single product announcement. It was a strategic reset delivered through a sequence of interconnected launches: a new class of AI models, a personal AI agent that runs in the cloud without any device needing to be open, a redesigned Search, a new subscription structure, and the clearest public statement yet that Google no longer sees AI as a smarter chatbot or a better search engine. It sees AI as an agent infrastructure. I/O 2026 is the moment it went public with that bet.

Some of what Google announced ships immediately. Some ships this summer. Some is months away. The breakdown below covers what's real and when.

Gemini 3.5 Flash and Spark: Google's Bet on Always-On Agents

Gemini 3.5 Flash is the model Google placed at the center of nearly every announcement. The company calls it the their strongest agentic and coding model yet, a model designed not just to answer questions but to plan and execute multi-step tasks, track state across those steps, and make decisions when something unexpected happens. That's a different design goal from retrieval-focused models. It's harder, and it's the capability that makes agents possible.

Gemini 3.5 Flash is now the default model for AI Mode in Google Search globally. That means any user anywhere accessing Google's conversational search is running on this model. The decision to deploy it as the default, rather than reserving it for subscribers, signals Google's confidence in its cost-to-capability ratio at scale.

Gemini Spark is the product built directly on top of Gemini 3.5 Flash, and it's the most ambitious thing Google announced at I/O. Spark is a personal AI agent that runs on virtual machines hosted in Google Cloud. It operates around the clock, 24 hours a day, without requiring a user's laptop or phone to be active. It's not a chatbot you open. It's a background process you authorize.

Google's demo showed Spark handling wedding planning: tracking RSVPs across email and calendar, sending follow-up reminders to guests who hadn't responded, coordinating schedule conflicts, and reporting status when asked. That use case is deliberately chosen. Wedding planning spans weeks, involves multiple external parties, requires tracking changing information, and degrades when any detail gets dropped. It's exactly the category of task where humans are inconsistent and AI agents are theoretically suited, not because the individual tasks are hard, but because sustaining attention across weeks is.

Spark uses Gemini 3.5 Flash alongside Antigravity, Google's agent-first development framework, to plan subtasks and execute them in sequence. Native integration with Google Workspace (Gmail, Calendar, Docs) is available at launch. MCP support for third-party apps is coming in the next few weeks. MCP, the Model Context Protocol, is the open standard that defines how AI agents connect to external tools and APIs. Support for MCP means Spark won't be limited to Google's own ecosystem long-term. Developers building products through MCP will be able to connect Spark to their own APIs once the integration ships.

Spark rolls out to trusted testers this week and to Google AI Ultra subscribers in the U.S. next week. Chrome support arrives later this summer. Only U.S. subscribers on the $100 or $200 AI Ultra plans will have access at launch.

One feature confirmed for later this year: Spark will be able to make payments on a user's behalf. The demo showed Spark authorized to purchase specific items within parameters the user set (brand, price ceiling, specific product type), completing the checkout without requiring any direct user action. This is the feature that moves Spark from useful assistant to authorized proxy. You set the rules. It acts within them.

Universal Cart, announced alongside Spark, uses AI to proactively monitor shopping carts and understand purchase context. If you're buying components for a PC build, it understands the assembly goal and can flag missing parts or incompatibilities. It's a narrower version of the same concept: AI that anticipates what you need rather than waiting to be asked.

The agent strategy doesn't sit in isolation from the broader industry contest over developer tooling. Google's I/O announcements show one side of that contest; our coverage of Anthropic's acquisition of Stainless, the SDK tool used by OpenAI, Google, and Cloudflare, covers another dimension of the infrastructure competition happening underneath agents.

What Changed in Search, Workspace, and the Gemini App

The biggest shift in Google Search isn't one feature. It's a reorientation of what Search is for.

The redesigned Search box does more than autocomplete queries. Google says the new interface anticipates user intent based on what you've looked for before, what you're in the middle of, and what logical next step follows from your current query. The implicit contract between user and search engine is changing: you're no longer querying a database. You're directing an agent.

Information agents in Search go further still. These are AI agents that run in the background monitoring specific topics, tracking changes over time, and alerting users proactively. Flight prices, sports scores, competitor news, topics you care about: the agent watches and surfaces updates. Information agents launch first for Google AI Pro and Ultra subscribers this summer, before wider rollout.

Agentic booking is also expanding. Google is extending its capability to find and book local experiences and services directly from Search, allowing users to specify detailed criteria and have the agent identify and complete matching bookings. The transactional layer in Search is becoming more capable.

Antigravity, Google's agentic coding framework, is coming to Search as generative UI. Search can now build custom interactive dashboards, graphs, simulations, and trackers in response to complex requests. If you're planning a home move, Search can build a checklist tracker with timelines. If you're following a developing story, it can build a dashboard that updates as new information arrives. This feature comes to all Search users this summer, with no subscription required.

The Gemini app received its own agent layer through Daily Brief, a feature that delivers scheduled proactive summaries of what's happening in areas the user cares about. Daily Brief handles regular check-ins; Spark handles long-running multi-day tasks. The two features are designed to work alongside each other, not to duplicate each other.

Workspace got voice capabilities across Gmail, Docs, and Keep. You can now dictate, edit, and organize documents and emails through voice interaction powered by Gemini. Google Pics is a new AI-assisted design tool built into the Workspace ecosystem. AI Inbox updates bring smarter prioritization and summarization to Gmail, reducing the time spent triaging rather than responding.

Android 17 was previewed at I/O with a feature called Android Halo, a synchronization layer for agents across multiple devices. If Spark is managing a task that involves your phone, laptop, and glasses simultaneously, Android Halo keeps the state consistent across all three. It's infrastructure for a world where agents are persistent processes spanning multiple surfaces.

The Android XR glasses make that physical. Google showed working demos of intelligent eyewear built in collaboration with Warby Parker and Gentle Monster, two premium consumer eyewear brands, rather than the developer-focused hardware that characterized earlier XR efforts. The glasses integrate Gemini deeply and respond to voice commands. They arrive this fall. The choice to partner with fashion brands, not just technology manufacturers, signals that Google is trying to build a product people will actually wear daily, not a demonstration device.

For developers and businesses deciding where to build, understanding how Google's models and agent infrastructure compare to OpenAI's and Anthropic's is increasingly consequential. Our guide comparing Claude, GPT, and Gemini covers model capabilities, pricing, and developer tooling across the major platforms.

Omni, Developer Tools, and the New $100 Ultra Plan

Gemini Omni is the model announcement that's easiest to underestimate and probably shouldn't be. It's a multimodal AI that reasons across text, images, audio, and video simultaneously, producing outputs that reflect understanding across all four modalities at once rather than processing them in sequence and stitching results together.

Google CEO Sundar Pichai described Omni as able to "create anything from any input." The more specific claim is world understanding: when you give Omni a combination of images, audio, and text, it reasons across all of them coherently, producing video output that reflects an understanding of physics, culture, history, and science. Earlier multimodal models tended to produce outputs that looked like independent components pasted together. Omni is built to avoid that.

Conversational editing is the consumer interface. You can change characters in a generated video, swap backgrounds, alter styling, or modify scene elements using plain voice commands. If you've worked with AI video tools before, you know that editing a specific element without regenerating the entire scene has been the hard problem. Conversational editing addresses this directly. The same applies to photos: you edit them with natural language rather than layer-based software.

The first model in the family, Gemini Omni Flash, arrives this summer. It will be available via API, which means developers can start building on top of Omni's multimodal capabilities before the feature reaches consumers broadly.

Google AI Studio received significant updates. Native Android vibe coding support lets developers build Android apps through conversational AI coding without leaving the development environment. New Workspace integrations connect AI Studio to Google's productivity suite. A mobile app for AI Studio launches alongside the other I/O announcements. The Gemini API gains enhanced capabilities supporting the new models and agentic workflows.

For researchers, Google announced Co-Scientist, a multi-agent system for scientific research assistance. Co-Scientist can help plan experiments, review literature, suggest hypotheses, and synthesize findings across domains. It targets a different user than Spark or Search's information agents: researchers and technical professionals managing long analytical workflows, not general consumers handling daily tasks.

Google Flow for AI-powered filmmaking received upgrades at I/O, as did Flow Music, Pomelli, and Stitch. Each targets a specific creative workflow: video production, music generation, interactive product experiences, and visual design. These tools represent Google's effort to establish an AI creative suite alongside its productivity and developer tools.

The subscription restructuring is the announcement with the most immediate impact for most users. According to Google's official subscription announcement, the new $100 AI Ultra plan includes a 5x higher usage limit than the $20 AI Pro tier, 20 terabytes of cloud storage, and YouTube Premium. The previous $250 plan remains available at $200, with the same capabilities as before.

Cutting the entry price for Ultra by 60% is a competitive signal as much as a product decision. OpenAI's and Anthropic's top-tier plans sit in the $100-$200 range. Google's previous $250 tier was structurally above those alternatives. The new $100 Ultra puts Google's most capable features into the same pricing band as its primary competitors, removing price as a reason to choose a different platform.

The $100 tier is also the required tier for early access to Gemini Spark. That's deliberate pricing architecture: the most ambitious product announcement at I/O gates behind the plan Google is now positioning as the starting point for serious users rather than the premium exception.

Taken together, the I/O 2026 announcements form a coherent argument that Google has been building toward for several years. Agents need capable models: Gemini 3.5 Flash. Agents need infrastructure to run continuously: Google Cloud virtual machines through Spark. Agents need ways to connect to external tools: MCP support. Agents need interfaces users will actually interact with: redesigned Search and the Gemini app. Agents eventually need to act in the world: the payments feature coming to Spark later this year.

Not everything announced is shipping today. Gemini Spark is in limited U.S. early access. Omni Flash arrives this summer. The payments feature for Spark is months away. Chrome support for Spark follows later this summer. These timelines matter because they define when the agent-first vision is available to use rather than just demonstrated on stage.

But the direction is no longer ambiguous, and the specificity of what was announced (real model names, real pricing, real availability dates) is different from the aspirational roadmaps that characterized earlier AI keynotes from every major company. Google I/O 2026 was a product conference, not a research preview.