1. Microsoft Will Let Copilot Access Local Files and Control Windows, With Features Rolling Out in the Coming Months Microsoft has demonstrated an expansion of Copilot that will allow the AI system to search local files on a Windows PC and perform actions across the operating system.

2. Anthropic Launches Haiku 5.5, Cutting Short-Context Prices to $0.10 and $0.50 per Million Tokens Anthropic has launched Claude Haiku 5.5, a small model aimed at high-volume, cost-sensitive work such as summaries, classification, database queries and coding subagents. The previous Haiku 4.

3. SafeActBench Finds Some AI Agents Exceed 95% in Static Judgment but Reach No More Than 52% in Interactive Execution AI agents that use tools can change external systems, but reaching the correct outcome does not prove that an action was justified by evidence gathered beforehand.

In Brief

• ChatGPT Adds Interactive Visuals and Tools to Responses OpenAI is rolling out an “Intelligent UI” that lets ChatGPT combine text with diagrams, charts, forms, buttons, and inline tools. The GPT-6-powered feature is launching globally for paid users before expanding to Go and free tiers.

• OpenAI Publishes 722 Manuscripts Claiming Hundreds of Math Results OpenAI released 722 manuscripts spanning 372 related result families, including claimed solutions to hundreds of open mathematical questions. Mathematicians are still assessing the work, which was produced by an unreleased frontier model and published with reasoning summaries and compute estimates.

• Meta and Microsoft Reportedly Curb Employees’ Claude Use Meta and Microsoft are reportedly steering employees toward their own coding tools while retaining Anthropic models in customer-facing services. Microsoft reduced many internal monthly AI spending ceilings, while Meta’s Claude Code user count reportedly fell from about 60,000 to 30,000.

• Google, Meta, and Isomorphic Labs Back a $300 Million Virtual-Cell Project Google DeepMind, Meta, and AI drug-discovery company Isomorphic Labs are jointly investing $300 million in Biohub’s effort to build a predictive virtual model of human cells. The work is part of a $1.8 billion initiative intended to let researchers investigate biological questions through digital simulations.

• Nous Research Raises $90 Million and Launches Enterprise AI Agents Open-source AI startup Nous Research raised a $90 million Series B at a $1.5 billion valuation and introduced Hermes for Businesses. The product lets companies deploy customized agents for multistep workflows while keeping their data private, according to Nous.

• Youth-Safety Group Calls ChatGPT for Teens an “Unacceptable Risk” Common Sense Media said its tests found unreliable parental alerts, inadequate crisis assistance, and continued homework completion in ChatGPT for Teens. OpenAI disputed the assessment’s methodology, while the nonprofit said it stands behind its findings.

• Google Opens SynthID Media Verification to Everyone Google launched a public website that checks images, video, and audio for SynthID watermarks added by supported AI-generation tools. Google says the verification system already handles one million requests daily, though watermark detectors can fail to identify some generated media.

• Meta Deploys AI to Find Ads Leading to Child-Abuse Material Meta introduced a language-model system that examines where apparently harmless ads direct users, aiming to uncover links to child sexual abuse material outside its platforms. The company is also deploying an AI red-teaming agent to probe its safety systems for weaknesses.

• Microsoft Unveils Nvidia-Powered Surface AI PCs and Agent Sandboxing Microsoft announced a Surface Laptop Ultra starting at $2,600 and a $6,000 Surface RTX Spark Dev Box, both designed to run AI models locally. A revamped Windows 11 also adds “Execution Containers” for sandboxing AI agents, which Microsoft says will be available to all Windows 11 users.

• Healthleap Raises $38 Million for Hospital Risk-Screening AI Healthleap raised $38 million for software that analyzes clinical notes, laboratory results, and vital signs to flag inpatients who may need further review. The company says the system does not diagnose patients and is deployed in more than 50 hospitals.

• Google Tests Prompt-Based Browser Game Creation Google Labs launched Playground, an experimental platform that generates 2D or 3D browser games from text prompts and uploaded visuals. It is initially available to adults in the US, with free generation limited by weekly tokens and higher allowances for Google One AI subscribers.

• TRACE Reports Up to 5.4× Faster Low-Precision AI Training Rollouts Researchers introduced TRACE, a quantization-aware method designed to align FP4 training and rollout execution for mixture-of-experts language models. Across four large models, they report performance comparable to BF16 rollouts with joint FP4 weights, activations, and KV caches, alongside rollout speedups of up to 5.4 times.

Read the full edition →