Tech • IA • Crypto
Google Just Dropped COSMO Then Mysteriously Pulled It
7/10TL;DR
A wave of rapid AI developments includes a briefly released Google assistant app, hints of a unified Gemini video model, healthcare AI advances, new developer tools, and rising competition in open-weight models.
KEY POINTS
Google’s “Cosmo” App Appears and Vanishes
Google briefly published an experimental Android app called Cosmo, then removed it within hours. The app combined local Gemini Nano, cloud processing, and a hybrid mode that dynamically switched between them. It also used Android accessibility APIs to read on-screen context, suggesting ambitions for a deeply integrated assistant. Early testing showed incomplete features and rough presentation, indicating it may have been released prematurely.
Hybrid AI Assistant Strategy Emerges
Cosmo’s architecture signals a broader shift toward on-device plus cloud AI systems. Lightweight tasks can run locally for speed and privacy, while complex queries are offloaded to servers. This hybrid approach could define future assistants by balancing performance, cost, and data sensitivity.
“Omni” Hints at Unified Gemini Media Model
Leaks point to a new Gemini “Omni” system tied to video generation. Unlike the current split between Veo (video) and Nano Banana (image models), Omni may unify media generation under a single framework. Its appearance in user-facing interfaces suggests an active rollout, potentially timed with Google I/O 2026. The move comes amid strong competition in AI video, where rivals are leading benchmarks.
DeepMind Tests AI Co-Clinician
Google DeepMind introduced a research system designed to assist doctors rather than replace them. The AI supports note-taking, research, and patient interaction under physician supervision. In testing, it produced zero critical errors in 97 out of 98 cases and outperformed existing medical tools in some scenarios.
Telemedicine with Vision and Voice AI
Built on Gemini and Project Astra, the system can see, hear, and respond during video consultations. It successfully guided patients through physical checks, such as inhaler use and shoulder mobility tests. However, human doctors still outperform AI in identifying serious conditions, reinforcing its role as a support tool.
Safety Measures in Medical AI
A dual-agent system adds oversight: one AI interacts with patients while another monitors for safety compliance. The system is currently limited to research trials across countries including the U.S., India, and Singapore, and is not approved for diagnosis or treatment.
OpenAI Codex Adds “Animated Pets” and Workflow Tools
OpenAI’s Codex introduced pixel-art “pets” that display task updates and allow interaction during processes. While playful, the feature doubles as a lightweight communication interface. Users can create custom pets via image uploads, fueling rapid community sharing.
Improved Developer Portability
Codex now imports configurations from other coding tools, reducing friction when switching environments. Additional updates include customizable voice dictation dictionaries and better workflow automation, positioning Codex as a broader AI desktop layer rather than a single-purpose coding tool.
Anthropic Prepares “Claude Jupiter”
Anthropic is reportedly red-teaming a new model codenamed Jupiter, following internal naming patterns used before major releases. The testing aligns with its safety-first deployment strategy and may precede announcements at an upcoming developer event. The update could expand the Claude 4.7 lineup or introduce a new generation.
Mistral Faces Backlash Over Medium 3.5
Mistral Medium 3.5, a 128B parameter model, launched with strong technical features and unified architecture. However, its pricing—$1.50 per million input tokens and $7.50 per million output tokens—drew criticism amid cheaper and competitive alternatives.
Rising Pressure from Open-Source Rivals
Models like Qwen 3.6 offer comparable performance at lower cost and with permissive licenses. Chinese open models increasingly dominate benchmarks and developer adoption, challenging Mistral’s position.
European Positioning Remains a Key Advantage
Despite criticism, Mistral retains strategic importance as a European, self-hostable, and GDPR-aligned option. This makes it attractive for regulated industries, with firms like HSBC already committing to deployments.
CONCLUSION
The AI landscape is rapidly fragmenting into hybrid assistants, multimodal systems, and regionally strategic models, with competition intensifying across both performance and deployment flexibility.