Headlines

1. OpenAI and Anthropic Announce Shared Agentic Safety Standard After White House Summit
OpenAI, Anthropic, Google DeepMind, and Microsoft jointly adopted a common “Agentic AI Runtime Safety” spec requiring toolbox sandboxing, human-in-the-loop kill switches, and real-time audit logging for autonomous agents. The deal follows a series of high-profile AI-agent-induced financial fraud incidents. Source: Reuters

2. EU AI Act Enforcement Begins in Earnest: First Major Fine Levied Over Deepfake Political Ads
The EU AI Office fined a large social platform €45 million for failing to label AI-generated political ads during German state elections, marking the first significant penalty under the AI Act’s transparency provisions. The decision is expected to set a precedent for model providers and advertisers across the bloc. Source: Bloomberg

Model Releases

3. Google DeepMind Releases Gemini 3.5 Pro with 10M-Token Context and Native Video Generation
Gemini 3.5 Pro arrives with a 10-million-token context window, live video understanding, and native 4K video generation, while maintaining a 3x efficiency gain over the previous generation. Early benchmarks show it outperforms current frontier models on Gauntlet 2.0 agentic tasks. Source: The Verge

4. Mistral AI Launches Codestral-2-OCR, an Open-Weight Multimodal Agent for Codebase Migration
The 52B-parameter model combines code reasoning with OCR and layout understanding, letting it refactor legacy codebases from scanned documentation and handwritten diagrams. Codestral-2-OCR is released under Apache 2.0 and reportedly matches closed competitors on SWE-Lancer for retrieval-heavy tasks. Source: TechCrunch

Industry

5. Microsoft and OpenAI Break Ground on $50B Phase Two of Stargate Datacenter Complex
Microsoft and OpenAI announced completion of the first phase and start of the second phase of the Stargate project in Cheyenne, Wyoming, adding 1.8 GW of compute capacity and a dedicated on-site nuclear microreactor pilot. The expansion is tied to training GPT-6-class models, expected in late 2027. Source: Wall Street Journal

6. Anthropic Launches Claude Conductor, an Enterprise Agent Orchestration Platform
Anthropic’s new platform offers multi-agent workflows, document-centric memory, and audit trails for regulated industries, priced at $100 per seat per month. It integrates with SAP and Salesforce, and has already been adopted by major accounting and healthcare firms. Source: CNBC

Research Highlights

7. “Scaling Test-Time Compute for Real-World Robot Manipulation”
A joint paper from UC Berkeley and Toyota Research Institute demonstrates that a 3B-parameter policy with expanded test-time search surpasses a 70B-parameter model on 34 dexterous manipulation tasks. The authors show that fine-grained physical simulation can serve as a world model for real robots. Source: arXiv:2610.03847

8. “Memory Layers Are All You Need: Sublinear-Parameter MoE at 10x Scale”
Meta FAIR researchers introduce “Memory Layers 2.0,” a sparse architecture that decouples storage from compute, achieving GPT-4-level perplexity on MedicalQA with only 20B active parameters. The paper proposes memory-augmented transformers as a cheaper alternative to standard mixture-of-experts. Source: arXiv:2610.02119

Compiled from multiple media reports.