Google Just Dropped COSMO Then Mysteriously Pulled It

Google just quietly dropped a mysterious AI app, then pulled it within hours. Gemini may be getting a new omni video model. DeepMind is testing an AI co-cl clinician for doctors and tele medicine. Open AAI added animated pets and smarter workflow tools to codeex. Anthropic is red teaming a new clawed build called Jupiter. And Mistrol's new open- source model is getting criticized for its price and performance. So, let's talk about it. All right. First, Google quietly pushed a new Android app to the Play Store called Cosmo. The listing described it as an experimental AI assistant application for Android devices. And the package name was com.google.ressearch.air.cosmo, which already made it feel more like a research test bed than a finished consumer product. The pitch sounded familiar. Cosmo was supposed to bring artificial intelligence directly onto your device, help organize your day, answer complex questions, and work behind the scenes to simplify your life. So, initially, it sounded almost like another Gemini app. The interesting part is what people found inside. Cosmo can run with a local Gemini Nano model, a remote PI server, or a hybrid mode that switches between local and server AI depending on what is available. That means Google may be testing a more flexible assistant setup where some tasks happen directly on the phone while heavier tasks go through the cloud. It also taps into Android's accessibility service API, which means it is designed to read or interact with what is happening on your screen. In theory, that is exactly the kind of access a serious phone assistant needs. It can see context, understand what app you are using, and potentially help across the whole device. In testing though, this feature did not seem fully ready. And that basically sums up Cosmo. It had specific AI skills, some enabled and some disabled. The assistant worked, yet it felt rougher than the main Gemini app. Even the Play Store screenshots were stretched into the wrong aspect ratio, which made the whole thing look like it went public too early. Then Google pulled it. On May 1st, after the app had been spotted, the listing disappeared for most users. people who had already installed it could still view the page when logged into the same account while everyone else got a not found message. So either Google accidentally published an internal experiment or it released something early and immediately realized it was not ready for public attention. And while that was happening another Google leak started pointing towards something much bigger, a possible new Gemini video generation tool called Omni. A screenshot from Gemini's video generation tab showed the line, "Start with an idea or try a template powered by Omni." That matters because Omni appears in the actual visible interface, not only buried in hidden code. Right now, Gemini's video generation flow is powered by VO3.1, while image generation is tied to Nano Banana 2 and Nano Banana Pro. Google describes Nano Banana Pro as built on Gemini 3, while Nano Banana 2 is tied to Gemini 3.1 flash image. So, the big question is whether Omni is just a new wrapper around VO, a new video model, or an early version of a larger Gemini Omni model that can handle images and video inside one system. That would be a major shift. Google currently has a split media strategy. VO handles video. Gemini based nano banana models handle images. Omni could bring those tracks closer together and make Gemini feel more like one unified creative system instead of a collection of separate models. The timing also makes sense. Google IO 2026 runs May 19th and May 20th and Google already said Gemini and broader AI updates will be part of the event. So if Omni is real, IO is the obvious stage and this is coming during a very competitive moment in AI video. Bite Dance's Cedance 2.0 Zero has been sitting at the top of video generation benchmarks. So Google has a clear reason to push harder here. Then we have Google DeepMind's most serious announcement from this batch, AI coinition. This is basically Google's idea for how AI could help doctors without replacing them. The problem is simple. Hospitals need better care, lower costs, and faster support. While the world is heading toward a shortage of more than 10 million health workers by 2030, Google's idea is a three-part care system. The patient, the doctor, and an AI assistant working under the doctor's control. So, the AI can help with research, medical notes, patient support, and live conversations while the physician still makes the real decisions. And the early results are interesting. Google tested the system on 98 realistic primary care questions. In 97 of them, the AI made zero critical errors. Doctors also preferred its answers over leading medical evidence tools. Then Google pushed it further with medication questions from Open FDA RXQA. These are tricky because medicine is full of details, side effects, interactions, and edge cases. On open-ended questions, the kind doctors actually ask in real life. Google says AI co-clin beat available frontier models. The most futuristic part was tele medicine. Google built on Gemini and Project Astra to give this AI eyes, ears, and a voice. In simulated video calls, it could listen, watch, and guide patients through basic physical checks. It corrected someone's inhaler technique and helped guide shoulder movements to check for a rotator cuff injury. Still, Google is being careful here. Human doctors performed better overall, especially when spotting serious warning signs and guiding critical exams. The AI matched or beat primary care doctors in 68 out of 140 tested areas, which is impressive, but it also proves the point. This is a helper, not a replacement. To make it safer, Google uses a dual agent setup. One AI talks to the patient while another AI watches the conversation and checks that it stays within safe medical limits. For now, this is research only. It is not meant to diagnose, treat, or give medical advice. Google is testing it with partners across the US, India, Australia, New Zealand, Singapore, and the UAE. Now, over at OpenAI, Codex got one of the weirdest updates we've seen from a serious coding product, animated pets. The Codeex desktop app now has pixel art pets that sit as overlays on top of the screen, even when Codeex is minimized. There are eight predefined pets, and they show little message bubbles about what Codex is doing in the background. If a pet speaks while a task is running, you can click it and reply back to the agent. So, what looks like a cute status indicator also becomes a small interaction channel. Users can summon or hide them with the / pet command. Then there is hatch, a bundled skill that lets users upload any image and turn it into an animated pet. The pet gets saved inside the local codec home folder so people can package and share them. Community directories like pet share and pet decks already started appearing and X filled up with custom creations almost immediately. The playful part is easy to mock. Yet the same update also adds more practical features. Codeex can now autodetect configuration files left behind by other coding agents including cloud codes cloudMD and import them. That means project rules, plugins, and conventions can move across tools with less manual setup. For developers, switching between agents because of weakly limits or different strengths that lowers friction. There's also a dictation dictionary in settings where users can preload abbreviations and phrases that voice input usually gets wrong. Put together, this makes codeex feel less like a simple coding assistant and more like a desktop layer for AI work. Open AAI is adding personality, portability, and voice polish around the agent experience. Raw coding performance still matters obviously yet stickiness and daily workflow are becoming part of the product race too. Anthropic may be preparing its own move. A new internal build called Claude Jupiter version 1 has reportedly entered red teaming ahead of the code with Claude developer conference in San Francisco on May 6th. The name Jupiter is probably just an internal code name. Anthropic has used planet names before for pre-release safety testing. Last year, a similar process used the code name Neptune before the Claude 4 family launched. That pattern is why people are paying attention now. The current Claude lineup has Opus 4.7 as the flagship. Sonnet 4.7 and Haiku 4.7 are still missing, which leaves room for a mid-tier and small tier refresh. There is also speculation that this could connect to the Mythos Foundation that surfaced in earlier reporting. The red team process itself fits Anthropic's responsible scaling policy, which includes jailbreak probes and constitutional classifier stress tests before Frontier deployments. So, the May 6th event could bring a full new generation, a Clawed 4.7 expansion, or something else. Either way, Jupiter suggests Anthropic has something close enough to test seriously. And finally, Mistl AI dropped Mistl Medium 3.5 and the internet reaction was rough. Mistl announced the model on April 29th. It is a dense 128 billion parameter model with agentic features. The release also included Mistral Vibe CLI which runs remote coding agents in the cloud, pushes pull requests to GitHub and can run tasks in parallel. Lehat also got work mode designed for multi-step autonomous tasks like email triage, research synthesis, and cross-tool workflows. On benchmarks, Medium 3.5 scores 77.6% on S.Bench to ebench verified which tests whether a model can fix real github issues with working patches. It also reaches 91.4% on tow cubed telecom a benchmark for agentic tool use in specialized environments. Mistral also merged medium 3.1 magistral and devstrol 2 into one unified set of weights with configurable reasoning effort. That is genuinely useful engineering. The problem is price and competition. Mistral charges $1.50 50 per million input tokens and $7.50 per million output tokens. Meanwhile, Alibaba's Quen 3.6 has 27 billion parameters, less than a quarter of Mistral's size, scores 72.4% on S.Bench verified, and ships under Apache 2.0, so developers can download it and run it freely. Chinese open source models like Quen GLM from Jepu AI and Mimo V2 from Xiaomi. now dominate much of the open source conversation. Mistrol Medium 3.5 has not even ranked on major independent leaderboards yet since thirdparty evaluations are still pending. That is why the reaction was so mixed. Critics argued that Mistrol is expensive, weaker than top rivals, and falling behind the Chinese open- source wave. Others pointed out the one thing that still makes Mistral matter. It is one of the only serious Western openweight options for European enterprises. That matters a lot. Banks, governments, and companies dealing with GDPR may avoid Chinese infrastructure and may also want alternatives to American AI labs. Mistral is EU headquartered, auditable, self-hostable, and legally easier for many European buyers. HSBC already signed a multi-year deal with Mistl to self-host models on its own infrastructure. All right, so the next few days could be very interesting, especially with Anthropics event on May 6th and Google IO coming later this month. Thanks for watching and catch you in the next one.