Google Just Dropped COSMO Then Mysteriously Pulled It

Key Concepts

• Cosmo: An experimental, short-lived Android AI assistant by Google.
• Omni: A potential unified video/image generation model for Gemini.
• AI Co-clinician: A Google DeepMind research project designed to assist, not replace, medical professionals.
• Codex (OpenAI): A coding assistant platform now incorporating "animated pets" and workflow interoperability.
• Claude Jupiter: An internal Anthropic build currently undergoing red-teaming.
• Mistral Medium 3.5: A new dense 128B parameter model facing criticism regarding pricing and competitive performance.

---

1. Google’s Experimental AI Initiatives

• Cosmo App: Google briefly released "Cosmo" on the Play Store, an experimental assistant designed to run via local Gemini Nano models, remote servers, or a hybrid of both. It utilized Android’s accessibility service API to provide context-aware assistance. The app was pulled shortly after discovery, likely due to its unfinished state.
• Gemini "Omni": Leaks suggest Google is developing an "Omni" model for video generation. This represents a potential shift from a fragmented media strategy (where video and image models are separate) to a unified creative system. This is expected to be a focal point at Google I/O (May 19–20).

2. DeepMind’s AI Co-clinician

• Objective: To address the projected global shortage of 10 million health workers by 2030 by providing a "co-pilot" for doctors.
• Methodology: A three-part system (patient, doctor, AI) where the AI handles documentation, research, and patient support under physician supervision.
• Performance: In testing with 98 primary care questions, the AI made zero critical errors. In telemedicine simulations, it successfully guided patients through physical checks (e.g., inhaler technique, rotator cuff assessments).
• Safety Framework: Google employs a "dual-agent" setup: one agent interacts with the patient, while a second agent monitors the conversation to ensure it remains within safe medical boundaries.
• Status: Research-only; not intended for diagnosis or treatment. Testing is underway in the US, India, Australia, New Zealand, Singapore, and the UAE.

3. OpenAI’s Codex Updates

• Animated Pets: A unique UI update where pixel-art pets act as status indicators and interaction channels for background tasks. Users can create custom pets via the "Hatch" feature.
• Workflow Interoperability: Codex can now import configuration files (e.g., cloud.md) from other agents like Claude Code, reducing friction for developers switching between tools.
• Voice Polish: A new dictation dictionary allows users to preload specific technical phrases to improve voice-to-code accuracy.

4. Anthropic’s "Jupiter" Build

• Red Teaming: Anthropic is currently red-teaming an internal build codenamed "Jupiter."
• Context: Following the naming convention of "Neptune" (which preceded the Claude 4 family), industry observers speculate this could lead to a new generation of models or an expansion of the Claude 4.7 lineup.
• Event: The findings are expected to be relevant to the "Code with Claude" developer conference on May 6th.

5. Mistral Medium 3.5 and Market Criticism

• Technical Specs: A dense 128B parameter model featuring agentic capabilities, including the "Mistral Vibe CLI" for autonomous GitHub pull requests and multi-step task execution.
• Benchmarks: Scores 77.6% on SWE-bench Verified and 91.4% on Tau cubed Telecom.
• The Controversy: Critics argue the model is overpriced ($1.50/input, $7.50/output per million tokens) compared to smaller, open-source alternatives like Alibaba’s Qwen 3.6, which offers competitive performance under an Apache 2.0 license.
• Strategic Value: Despite performance criticism, Mistral remains a critical "Western open-weight" option. Its EU headquarters and self-hostable nature make it a preferred choice for European enterprises (e.g., HSBC) subject to strict GDPR and data sovereignty regulations.

---

Synthesis

The current AI landscape is characterized by a shift from raw model performance to "workflow stickiness" and specialized application. While Google and Anthropic are preparing for major product reveals, the industry is seeing a clear divide between high-cost, proprietary Western models and highly efficient, open-source models from international competitors. The focus has moved toward integrating AI into daily professional workflows—whether in medicine, coding, or enterprise infrastructure—while balancing the need for safety, regulatory compliance, and user engagement.