Within hours, it was downloaded, reverse-engineered, and dissected by developers and journalists across the world. By the time Google pulled the listing, it was already too late. Two weeks before Google I/O 2026, the company had accidentally shown its hand.

The package, listed under com.google.research.air.cosmo, came from Google Research's AIR group, the AI Research division that works on next-generation assistant prototypes well before they reach consumers. The timing, the size, and the sophistication of what was inside it all pointed in one direction. This wasn't a small experiment. This was the future of Google's AI assistant strategy, accidentally made public two weeks too early.

What COSMO actually Is

COSMO is not simply a chatbot, it appears to function as a deeply integrated AI operating layer capable of helping users proactively unlike traditional assistants that wait for commands, COSMO is designed to anticipate user needs.

The architecture behind it is what makes it technically significant. While the current Gemini app relies heavily on cloud-based processing, COSMO is designed as an on-device AI agent. The core of COSMO is the Gemini Nano model, which allows the AI to function locally on an Android device without requiring an active internet connection. The app features a hybrid mode that intelligently switches between local processing using Nano and high-power server processing based on the complexity of the task and connectivity.
By tapping into Android's AccessibilityService API, COSMO can see what is on the user's screen to provide real-time assistance, giving it extreme context awareness that goes well beyond anything the current Gemini assistant offers.

Skills signal where Google is heading

What COSMO can do is as revealing as how it does it. The app is equipped with a suite of skills that allow it to act proactively based on user context. A calendar event suggester detects when the user agrees on a time during a conversation and offers to schedule it automatically. A quick photo lookup finds a photo the user mentioned wanting to share without requiring a manual gallery search. A document writer drafts letters and summarizes text when the user mentions needing a document. Deep Research compiles full reports from multiple sources for complex queries. A Conversation Summary provides a brief overview of recently ended discussions when the user switches tasks.

One of the most significant capabilities is support for Mariner browser automation an internal Google AI system capable of navigating websites and completing actions automatically. This would allow COSMO to go beyond answering questions and instead complete tasks directly for users.

Accidental release, deliberate direction

Despite its advanced capabilities, the COSMO app feels decidedly unfinished. The user interface is a very basic chat interface, and the Play Store listing appeared poorly executed, with screenshots squished into incorrect aspect ratios. The general consensus among tech analysts is that this was not meant for consumers it was a premature publication ahead of the official unveiling expected at Google I/O 2026.

The architectural pattern COSMO reveals a three-mode system combining Gemini Nano on-device processing, a PI server for cloud processing, and a hybrid routing layer is already standardising across the AI assistant industry. Apple has moved in this direction. Google is now clearly following the same path.

Whether COSMO launches as a standalone app, gets folded into Gemini, or arrives as something else entirely at Google I/O remains to be seen. What the accidental leak confirmed is that Google's next AI assistant will sit on the device, watch the screen, anticipate actions before being asked, and complete tasks rather than just answer questions.

Google I/O 2026 is scheduled for mid-May. The company now walks in knowing exactly what questions it will be asked because it accidentally answered most of them a fortnight early.

Read more news like this on www.etnownews.com