Home / News / Google accidentally published Cosmo, its new AI assistant, and deleted it hours later

Google accidentally published Cosmo, its new AI assistant, and deleted it hours later

google-accidentally-published-cosmo,-its-new-ai-assistant,-and-deleted-it-hours-later

Google had a monumental slip-up this week. The Mountain See company accidentally published a completely new application on Play Retailer called Cosmoan artificial intelligence assistant designed to live right on your Android device. The app was available for a few hours before Google hastily removed it, as if someone had pressed the wrong button at the wrong time.

The curious thing is that the leak was not a hack or a leak of information from within. Simply put, someone at Google published the app ahead of time. And of course, in a matter of minutes, the technology community had already discovered, downloaded and analyzed it with a magnifying glass.

Cosmo, the AI ​​assistant that Google didn’t want you to see yet

Cosmo is not just any assistant. Unlike Gemini, which works primarily in the cloud and requires an Internet connection for almost everything, Cosmo is designed to operate directly from the device, thanks to a model Gemini Nano integrated that can work even offline.

That makes it something much deeper and inner most. Basically, Cosmo would install itself on your phone and start listening, observing, and understanding the context of everything you do. It sounds a bit like science fiction, but Google already had mapped out several of the functions it planned to offer with this app. Among the capabilities that could be seen before the token disappeared from the store, things like:

  • Checklist Trackerwhich automatically suggests lists based on your conversations
  • A document writer which is activated when it detects that you mention needing to write something
  • A calendar event suggestercapable of scheduling meetings if it detects that you have made an appointment with someone
  • A browser agent that can automate tasks on the web using Mariner
  • The function of Deep Overviewdesigned for complex investigations that require multiple sources

In addition, Google had contemplated three modes of operation for the assistant. The mode Hybrid I would use cloud servers when there was net and switch to the local model when there wasn’t. The mode PI Supreme would depend exclusively on remote servers, while the mode Nano Supreme would work completely on the device. A well-thought-out architecture for an app that was supposedly just an internal experiment.

A leak that arrives just before Google I/O

The timing of the error could not be more revealing. Google I/O 2026 takes place in a few weeksand it is precisely at that event where the company usually presents its biggest innovations in artificial intelligence and instruments. Everything indicates that Cosmo was saved for that scenario, and that someone pressed “publish” much earlier than expected.

Users who managed to download it before it disappeared reported that the app was clearly in a very early stage of development. It worked in a rudimentary way, with many “processing” messages and an experience that left a lot to be desired. A user who accessed it using a VPN commented that the app bombarded with follow-up questions and did not respond correctlyconfirming that it still needed work before being ready for the comparable public earlier.

This reinforces the hypothesis that Cosmo was part of a internal or trusted tester programand that its appearance in the Play Retailer was an accidental launch that no one within Google had planned. The withdrawal of the app was equally fast: in a matter of hours, the link stopped working and the listing disappeared from the store.

What Cosmo reveals about Google’s AI plans

Beyond the stumble, what this incident makes clear is that Google is working on something much more ambitious than Gemini. Cosmo represents a new generation of assistants that not only answer questions, but live with the user, learn from their context and act proactively before they are asked. It is the leap from a reactive assistant to a truly autonomous one.

The difference compared to what we know today is enormous. Gemini can answer a query if you ask him. Cosmo, on the other hand, would notice that you are planning something and would act before you ask. He would listen to your conversation, detect that you mentioned a meeting at 3 PM, and offer to schedule the event without you asking. That level of contextual integration is what Google appears to be building with this project.

And although the app was removed, what users who explored it saw suggest that the product is more advanced than Google wanted to admit. The functions described in the sheet, the three operating modes and the integration with tools such as the calendar, the browser and the clock are not the outline of an idea. They are the skeleton of a product that already has shape.

Google has not officially commented on Cosmo. But with Google I/O just around the corner, chances are you won’t have to wait long to reveal what exactly slipped through your fingers this week. The inadvertent mistake ended up being, unintentionally, the most effective teaser that Google could have released.

Keep reading:
• Google launched an app that converts your voice into text without net and it is exactly what you needed
• Google and the Pentagon reach an agreement so that their AI has access to self-discipline classified material
• Google Translator turns 20 and comes loaded with AI functions that will change how you learn languages