In recent years, the presence of digital assistants on Android mobiles It has gone from being a novelty to becoming an integral part of everyday life. Asking questions by voice, creating reminders, or requesting that a message be sent is now quite commonplace, and yet the general feeling is that technology still has room to go one step further.
Within this context, the emergence of COSMO for AndroidAn experimental Google app, released briefly, offered a glimpse into the company's commitment to more contextual and proactive assistants. The app was mistakenly published on Google Play, available for a very limited time, and quickly removed, but it was enough for the community to analyze in detail what this new experiment proposed.
What is COSMO for Android and how did it get to the Play Store
COSMO was described in its Google Play listing as a “experimental AI assistant application for Android devices”It appeared linked to the identifier com.google.research.air.cosmo, which clearly points to a development originating from Google Research, but it was published from the company's main account in the app store.
According to the official text, its purpose was “bringing the power of artificial intelligence directly to your device”The idea was not simply to offer another chatbot, but a tool capable of moving behind the scenes, interpreting conversations, detecting needs, and proactively proposing actions, from organizing the day to answering complex questions.
COSMO's presence on the Play Store was very brief. Several specialized sources, such as 9to5Google, detected the app, downloaded it, and tested it, but it disappeared within hours. Google removed it from the storeThe company itself confirmed internally that it was a premature publication, which fits with the fact that development was still in a fairly early stage.
These kinds of fleeting appearances are not new in the sector, but in this case the timing was particularly striking: the alleged error occurred just weeks before the celebration of Google I/O 2026, the annual conference for developers where many of the new features in Android and artificial intelligence are usually announced.
An assistant designed to live inside your mobile phone.
One of the key aspects of COSMO for Android is its approach as assistant residing on the device itselfwith the ability to function even without an internet connection in many cases. Unlike other services focused almost entirely on the cloud, the app included a locally integrated Gemini Nano model.
That approach was already evident in the download size itself: around 1,13 GBThis is a high figure for an assistant app, but logical considering that the AI model had to be embedded in the phone. After installation, COSMO requested access to several system permissions to operate with greater context: screen, photos, contacts, and other elements typically used when an assistant needs to see what's happening on the device.
Upon opening it for the first time, users encountered a very simple chat interfaceIt was practically unadorned, clearly demonstrating its experimental nature. There were no grand visual flourishes or advanced configuration options beyond a few technical tweaks; everything was geared towards testing the model's behavior and capabilities.
The testimonies of those who managed to use it during those initial hours agree that its operation was still quite rudimentaryThe messages "processing", incomplete answers and constant follow-up questions were frequent, a pretty clear sign that the system was not yet ready for a mass release to the general public.
Main features: more than a generic chatbot
Beyond its unfinished appearance, what was truly relevant about COSMO were the features that Google had documented in the app's information sheetIt detailed a series of "Skills" designed to go far beyond simply answering a written or dictated question.
Among these capabilities were several tools geared towards everyday tasks, which clearly illustrate the project's philosophy:
- List tracker: suggested keeping or creating lists (for example, shopping lists or to-do lists) when items that fit that format were mentioned in a conversation.
- Document writerIf the user mentioned that they needed to write a letter, draft a text, or summarize information, COSMO offered to generate the document automatically.
- Suggested events on the calendar: upon detecting that a get-together, a meeting or any plan with a date and time was being discussed, it suggested creating an event in the calendar.
- Add timersIf a task with a time limit appeared in the conversation, the assistant suggested setting a timer in the clock app.
Alongside these more "domestic" functions, COSMO also offered tools related to navigation and information retrievalIt stood out. browser agent designed to automate tasks on the web through a component called Mariner, as well as the ability to launch quick searches when a question could be answered in seconds with a Google query.
Another particularly striking piece was the option called Deep ResearchThis function was geared towards complex investigations that require gathering information from multiple sources, combining that data, and returning a more detailed report. The goal was to distinguish between a simple question, which can be answered with a quick search, and an in-depth query, which requires more elaborate analysis.
In addition, the skills list included sections such as explaining technical jargon or acronyms, offering additional information when the user showed curiosity about a topic, and the ability to provide context about people and eventsAll of this is with the idea that the assistant better understands what is being talked about at any given moment.
Access to device context and conversation memory
One of the clear differences between COSMO and other more traditional attendees lies in its relationship with the content of the phone itselfAmong the documented functions was the "quick photo search", designed to locate specific images from a simple reference in the conversation.
For example, if someone said, “the photo I took on my last trip to the beach,” COSMO could offer to directly find that image to share without the user having to manually search through the gallery. This type of interaction relies on the model's ability to relate Natural language comments with elements stored on the device.
Similarly, the assistant was designed to summarize recent conversations When changing topics or resuming a previous thread, the system could remind the user what had been discussed in a chat, what the key points were, or what tasks remained pending—especially useful when switching between multiple conversations or projects.
Also included were functions under generic names such as "Remember", intended to to help the user retrieve information they are trying to rememberFrom specific data points to decisions made in a previous discussion, this creates an assistant less focused on direct commands and more on following the flow of what's happening around them.
This approach, of course, raises questions about privacy and data management, especially in Europe, where the regulatory framework It is more demanding. Although COSMO was clearly geared towards leveraging the device's context, the weight of Gemini Nano within the phone itself suggests that some of that processing could be done without sending all the data to the cloud, something that better aligns with the requirements of the GDPR and European regulations.
Gemini Nano and hybrid operating modes
From a technical point of view, one of the most interesting elements was the integration of Gemini Nano as a local modelThis is Google's AI model designed specifically to work on devices with limited resources, such as smartphones, allowing it to perform certain tasks without relying entirely on remote servers.
In the COSMO settings, there was a menu called “Fulfillment Model” with three different operating options:
- HybridIt combined the local model with cloud processing, using one or the other depending on whether a connection was available or the complexity of the task.
- PI OnlyIt depended solely on the model hosted on the remote server (the app identified it as PI), dispensing with the local component.
- Nano Only: it was committed to resolving as much as possible within the phone itself, using only the Gemini Nano.
It's not entirely clear what the label exactly means. "PI" In this context, however, everything indicates that it refers to a more powerful model running on Google's infrastructure. The existence of these three modes suggests that the application functions as test bench to measure the extent to which tasks can be delegated to the device and when it is necessary to resort to the cloud.
The settings also included options such as voice matching and the ability to allow access to what is displayed on the screenThis fits with the functions of summarizing conversations, assisting in navigation, and automating tasks through the browser agent.
For manufacturers and developers in Europe, an experiment of this kind can serve as a preview of How will generative AI models be integrated into the next generation of Android phones?without relying exclusively on permanent connections and allowing adaptation to local regulations.
A clearly premature launch
Everything points to the fact that the publication of COSMO on Google Play was a internal slip-up rather than a planned announcementThe combination of an identifier linked to Google Research, its presence in the company's main account, and the almost immediate removal fit with the idea of a premature launch.
Specialized media and users who managed to install the app describe a still very green experienceThere were numerous bugs, vague answers, and behavior more reminiscent of a tech demo than a product ready to reach millions of people. Some even accessed the app using a VPN and found that many requests were left unresolved or resulted in further follow-up questions, without ever reaching a clear solution.
This embryonic state reinforces the hypothesis that COSMO would be part of a internal testing programPerhaps with trusted employees or testers, and the "publish" button was pressed in the wrong place. This isn't the first time something like this has happened in the Android ecosystem, but in this case, the situation is more sensitive because it involves a project tied to Google's artificial intelligence strategy.
The speed with which the Play Store listing disappeared is also noteworthy. In a matter of hours, the The link is no longer operational. And the app was no longer accessible in a conventional way, although the prior analysis of the community left a good number of screenshots, descriptions and testimonials about its behavior.
Without detailed official statements from Google, the most widespread interpretation is that COSMO was reserved for to be presented in a more prepared context, probably during Google I/O, where the company usually frames its releases within a broader narrative about the evolution of Android and its services.
What COSMO hints about the future of AI on Android
Beyond the anecdote of the failed launch, what's relevant is what COSMO suggests about Where is Google headed with assistants on Android?The project moves away from the purely reactive approach of previous generations and aims for a model in which the assistant lives alongside the user, interprets the context, and acts before an explicit command is given.
The difference compared to simply having a chatbot on the screen is considerable, as demonstrated by the Opening Siri to third-party assistantsWhile a system like Gemini, in its most basic form, responds when consulted, COSMO is designed to detect patterns in what the user does and saysIf a meeting is mentioned, suggest scheduling it; if a recipe with specific times is discussed, suggest a timer; if technical terms appear, offer a quick explanation.
This type of behavior is reminiscent of the idea of a more autonomous digital agentIt's capable of making small decisions in the background to reduce friction in everyday mobile phone use. The key, if it ultimately materializes into a product aimed at the general public in Europe and Spain, will be how this level of integration is communicated and what controls are offered over the amount of data the assistant can access.
Another interesting aspect is the combination of local and remote modelsThe fact that COSMO allows you to choose between hybrid modes, server-only or local-only model opens the door to more flexible configurations, where the user or the system itself can prioritize speed, privacy or computing power depending on the situation.
For the Android ecosystem in Europe, where phones of very different price ranges and networks with varying quality coexist, this approach could be especially useful. A mid-range phone, for example, could rely more on the cloud, while a high-end device in Spain or other EU countries could better leverage the capabilities of Gemini Nano by running it directly on the device.
Overall, COSMO hints at a scenario in which the Assistants integrate more naturally with the operating systemThey turn routine actions into almost automatic processes and combine on-device intelligence with remote support when needed.
The brief appearance of COSMO for Android on Google Play was undoubtedly a setback, but it also served as a discreet window into the experiments Google is conducting to redefine how users interact with their mobile devices. If everything we've seen so far culminates in a final product, the company's next steps in AI assistants could mark a significant shift in how we, in Spain and the rest of Europe, use Android on a daily basis.
