Author: AINF

This week shows AI moving on two fronts at once: toward more powerful frontier systems, and toward real-world deployment at global scale. Anthropic introduced Claude Fable 5 and the restricted-access Claude Mythos 5, signaling a new model for releasing advanced capabilities with stronger safeguards. Apple expanded Private Cloud Compute to Google Cloud, using NVIDIA GPUs for Apple Intelligence workloads while extending its privacy architecture beyond its own data centers. Google brought Gemini 3.5 Live Translate into products and developer tools, making real-time speech translation across 70+ languages feel closer to everyday use. At the same time, Cognition’s FrontierCode benchmark reframed AI coding around mergeability rather than simple correctness, while Xiaomi pushed trillion-parameter inference past 1,000 tokens per second. And with ResNet receiving the CVPR 2026 Longuet-Higgins Prize, the week also offered a reminder that today’s AI breakthroughs are still built on research ideas that reshaped the field a decade ago.

This week shows how quickly AI is moving from frontier research into tools, devices, and physical systems people can actually use. Anthropic’s latest analysis puts recursive self-improvement back in focus, with Claude now writing more than 80% of the code merged into its production systems. That progress is already reaching users: OpenAI extended Codex with Sites, turning plain-language prompts into interactive websites and apps, while Google released Gemma 4 12B, a multimodal model small enough to run locally on a laptop. On Windows, Microsoft’s OpenClaw and NVIDIA RTX Spark point to a new layer of local-agent infrastructure, combining secure execution, enterprise control, and hardware purpose-built for running AI agents on personal machines. And with Unitree’s H2 Plus, the story moves into robotics, as AI begins to step out of software workflows and into machines that move.