Claude announced two AI models, Fable 5.1 and Mythos 5.1, pitching them as the world’s most advanced options for coding and knowledge work.

World Labs unveiled Atlas, which it calls the first AI system that creates images and video frames with precise camera control and rebuilds scenes in 3D, letting users model a scene, move the camera, and simulate space and time.

Google announced Gemini 3.8, calling it its best model for reasoning and coding, three weeks after 3.7 Flash, and introduced two variants: 3.8 Flash and 3.8 Flash Cyber. It uses long-running agent loops, letting the system take successive steps.

Meta released Muse Spark 1.3, which handles multi-step tasks and coding better, keeps several tasks going, asks clarifying questions, flags when stuck, confirms actions, makes up answers less, and needs ~20% fewer tool uses and ~25% less text.

Nvidia agreed to buy Hugging Face and says it will invest in the site’s infrastructure and growth, while keeping it a neutral home for open-source AI tools that work with any cloud or hardware.

Anthropic’s ClaudeDevs open-sourced Claude Commerce Agents, a blueprint and sample apps for building shopping and merchant assistants that handle tasks end to end, with working examples for retail, travel, telecom, and entertainment.

OpenAI’s Astra uses a “recurrent depth” reasoning method — letting the model loop through its own steps multiple times — which can boost performance and lower cost, but researchers warn it hides how the model reasons, complicating oversight.

Thomson Reuters launched a large language model called Thomson, trained continually on 175 years of its proprietary legal, tax, accounting, and news data, and says it spent $40 million building it.

U.S. Census surveys show AI adoption rose from 3.7% (2023) to ~10% by late 2025 (18% with a broader question), yet about 96% of firms report no employment change; AI mostly augments tasks like writing, search, and summarization.