MCP App StoreLe meilleur de l'écosystème IA et MCP, sélectionné chaque jour.
This week was defined by a clear shift from raw capability to operational maturity, with a heavy emphasis on standardization, economic efficiency, and the intersection of AI with physical robotics.
Standardizing Agentic Infrastructure The most significant structural update was the release of the MCP 2026-07-28 specification. By moving toward a stateless core and introducing Multi Round-Trip Requests, the Model Context Protocol is hardening authorization and scaling tool integrations—a move immediately embraced by Anthropic in Claude. Parallel to this, Google expanded its Managed Agents in the Gemini API, providing developers with the hooks necessary for production-ready autonomous systems.
The Efficiency Frontier OpenAI continues to optimize the "intelligence per dollar" ratio. The release of GPT-5.6 focuses on price-performance frontiers for Luna and Terra models, while research into ARC-AGI-3 benchmarks revealed that strategic inference-time settings (reasoning and compaction) can triple performance without changing the underlying model. This suggests that the next phase of LLM gains will come as much from orchestration as from training.
Embodied AI and Scientific Discovery AI is increasingly leaving the screen. Google DeepMind's Gemini Robotics ER 2 and NVIDIA's Cosmos-H-Dreams are pushing the boundaries of video understanding and real-time generative simulation for surgical robotics. Meanwhile, the application of coding agents to genomics and the launch of OlmoEarth for planetary-scale geospatial inference demonstrate that agentic AI is now a primary driver in scientific computing.
Key Stories:
OpenAI has released new findings on long-standing open problems in geometry, cryptography, and complexity. These advances push the boundaries of theoretical computer science and mathematical reasoning.
Google DeepMind introduces Gemini Robotics ER 2, showcasing significant improvements in video understanding and multi-robot collaboration. The model enhances tool orchestration for complex robotic applications.
OpenAI releases GPT-5.6 with improved price-performance ratios for the Luna and Terra models. These efficiencies are designed to help enterprises scale AI workflows more affordably.
Avatarin deployed a multilingual retail agent using GPT-Realtime for Yamada Denki, serving 30,000 users in two weeks. The implementation demonstrates the high efficacy of real-time voice agents in consumer retail.
OpenAI demonstrates how specific API settings for reasoning and compaction tripled GPT-5.6 performance on the ARC-AGI-3 benchmark. This highlights a critical optimization path for improving model efficiency and reasoning capabilities in complex problem-solving tasks.
OpenAI introduces GPT-5.6, focusing on maximizing the ratio of intelligence to compute. The update optimizes inference and agentic workflows to deliver higher utility per dollar for developers.
OpenAI explores how AI coding agents are modernizing scientific computing. The report highlights significant accelerations in software development and discovery within the genomics field.
Ai2 has launched OlmoEarth, a platform designed for planetary-scale geospatial inference. This infrastructure allows for more sophisticated AI-driven analysis of Earth-observation data.
Google is expanding its Managed Agents in the Gemini API, introducing support for 3.6 Flash and new hooks. These updates aim to help developers build more reliable, production-ready autonomous agents.
Liquid AI introduces LFM2.5-Encoders, optimizing long-context inference for CPU environments. This development significantly reduces the hardware barrier for running high-capacity context models.
The Model Context Protocol has updated to a stateless core, introducing Multi Round-Trip Requests and header-based routing. This update hardens authorization and provides a formal extensions framework for better scalability.
Cursor has launched a new 'Cursor Start' plan specifically for developers in India. This initiative aims to make high-end AI coding tools more accessible in the region.
Claude is integrating the latest MCP 2026-07-28 specification, moving toward a stateless core with standardized extensions. This rollout will enhance how Claude products interact with external data and tools.
NVIDIA introduces Cosmos-H-Dreams, a real-time generative simulation framework designed for surgical robotics. This advancement enables higher precision and safer training environments for complex medical procedures using generative AI.
A detailed technical post-mortem of a security breach involving AI agents at a frontier lab. It provides a critical timeline and analysis of the intrusion, offering essential lessons for securing agentic systems.