Google I/O 2026 opens on Monday, May 19 in Mountain View. By the time Sundar Pichai takes the stage, most of the surprises will already be public. In the past two weeks, tinkerers, leakers, and code diggers have pulled three major products out of the Gemini app before Google could announce them: an AI agent called Spark, a new model called Gemini 3.2 Flash, and a video generation system called Gemini Omni.

Each leak, on its own, would be significant. Together, they outline a company preparing to fight on three fronts simultaneously: agents, models, and media generation. Here is what surfaced, what it means, and why Google probably wanted to control the narrative.

Gemini Spark: An Agent That Lives in Your Chat

Spark is the most aggressive of the three leaks. First spotted via onboarding screens on May 14 and then activated by multiple tinkerers on X by May 15, Spark appears to be a persistent AI agent that lives inside the Gemini app and performs multi-step tasks across Google Workspace apps without human intervention.

According to Android Authority, which compiled reports from @Waguri_Kaoruko8 and @testingcatalog on X, Spark can clean junk from your Gmail inbox, assemble notes before key meetings, create custom news digests, and build reusable "skills" for recurring workflows. Users will reportedly be able to let Spark function without a human reviewing its work. It can also control the Chrome browser as an agent and use files stored on your computer.

The direct competitor is Anthropic's Claude Cowork, which launched earlier this year and lets Claude operate as a persistent desktop agent. Android Authority's headline was blunt: Spark "looks like it's gunning for Claude Cowork's throne." But there is a critical difference. Spark lives inside Google's ecosystem, which means it has native access to Gmail, Docs, Drive, Calendar, and Chrome. Claude Cowork has to reach those services through integrations. Google's version starts closer to the data.

Spark may also rely on a separate AI model for its functioning, which suggests Google could be deploying a purpose-built agent model rather than running Spark on the same general-purpose Gemini architecture. Third-party app support was not visible in the leaked screenshots, but given Google's history with Android intents and ecosystem lock-in, it would be surprising if it stayed Workspace-only.

Gemini 3.2 Flash: Sub-200ms Responses at $0.25 per Million Tokens

The second leak is a model, not a product. Gemini 3.2 Flash is reportedly optimized for speed above everything else. The core claim: many prompts return responses in under 200 milliseconds. That is fast enough to feel instantaneous in voice conversations, coding copilots, and real-time search.

According to NPowerUser, which analyzed the leaked specifications, Flash 3.2 performs surprisingly close to Gemini 3.1 Pro in general reasoning, summarization, coding assistance, and conversational responses. Google is reportedly using advanced distillation, sparse architecture, and improved routing systems to compress high-end capabilities into a much smaller and cheaper model.

The pricing numbers are what caught the developer community's attention: $0.25 per million input tokens and $2 per million output tokens. If those numbers hold, Gemini Flash would become one of the most cost-effective frontier AI models available, undercutting most competitors by a significant margin. The knowledge cutoff is reportedly January 2026.

Sources suggest the model could launch during I/O or potentially one to two days before the keynote, a pattern Google has used before to build pre-event momentum. There are also reports that Google may rebrand it to Gemini 3.5 Flash before launch, positioning it as a bigger leap.

Gemini Omni: A Video Generation Model Nobody Expected

The third leak is the wildest. On May 2, TestingCatalog spotted a UI string inside the Gemini app's video generation tab reading "Powered by Omni." Nine days later, on May 11, at least one Gemini AI Pro subscriber had access to the model and shared two generated clips: a spaghetti scene at a seaside restaurant and a professor writing trigonometric proofs on a chalkboard.

The spaghetti scene is a deliberate reference to the "Will Smith eating spaghetti" benchmark that early video models notoriously failed. According to Chrome Unboxed, Omni's version was "incredibly realistic," with convincing physical interaction between actors and food, no limb morphing, and proper handling of spaghetti strands as they were lifted and twirled. This is the kind of physical-object scene that broke Sora 1 in 2024 and that Veo 3 only started handling reliably in late 2025.

Nobody is sure what Omni actually is. There are three theories: a rebrand of the existing Veo 3.1 pipeline, a completely new in-house Gemini video model, or a true unified omni-model that handles text, image, video, and audio in one architecture. The UI positioning next to Google's internal "Toucan" codename (used for Veo 3.1) suggests a pipeline swap rather than a parallel offering.

One user burned 86% of their daily AI Pro quota generating just two clips, which hints at the compute cost. The model supports video remixing and editing directly in chat. If Google announces Omni at I/O, it would position Gemini as a direct competitor to OpenAI's Sora 2 and ByteDance's Seedance 2 in the video generation space.

Why This Matters

The leaks paint a picture of Google preparing a comprehensive offensive. Spark takes on Claude Cowork and OpenAI's custom GPTs in the agent layer. Flash 3.2 attacks the cost-and-speed layer where most real developer adoption happens. Omni targets the media generation market that OpenAI's Sora and ByteDance's Seedance currently dominate.

For Google, the timing is both strategic and unfortunate. Strategic because Monday's keynote was supposed to be a controlled reveal of each product with polished demos and rehearsed applause lines. Unfortunate because leakers have been pulling the wrapping paper off for two weeks. When Pichai walks on stage, the audience will already know the headlines. The question is whether the actual announcements are bigger than what leaked, or whether Google just lost control of its biggest event of the year.

Google I/O 2026 kicks off Monday, May 19 at 10 AM PT.