Microsoft, China release new models; agents proliferate

Anthropic / Claude ecosystem

No significant new developments.

Frontier model providers

Google DeepMind showcases generative AI at Tribeca with Dear Upstairs Neighbors | Let's Data Science

Google DeepMind presented 'Dear Upstairs Neighbors' at the Tribeca Film Festival, showcasing a hybrid generative-video production workflow. This project combined fine-tuned Veo and Imagen models with traditional animation techniques, demonstrating how custom model fine-tuning and video-to-video pipelines can enhance animator control.

Microsoft Declares "AI Independence" with In-House Reasoning and Coding Models | Wow News

Microsoft has launched its own proprietary AI model family, MAI (Microsoft AI), which includes MAI-Thinking-1, MAI-Code-1, MAI-Image-2.5, MAI-Voice-2, and MAI-Transcribe-1.5. This strategic move aims to reduce the company's dependency on OpenAI and cut operational costs by 90%, signaling a shift towards building sovereign intelligence stacks.

GLM 5.2: China's Open Frontier Model Dropped the Day Anthropic Got Banned [2026] - DEV Community

ZhipuAI has released GLM 5.2, a fully open-source frontier model boasting 744 billion parameters with MIT-licensed weights. This release occurred on the same day the US government restricted Anthropic's Claude Fable 5, positioning open-source Chinese AI as an alternative to restricted US closed-source models.

AI developer tooling & infrastructure

How to Build a QwenPaw Agent Workspace with Custom Skills, Model Providers, Console Access, and Streaming API Testing - MarkTechPost

A new tutorial demonstrates how to construct and deploy QwenPaw agent workspaces, including the implementation of custom skills, authentication mechanisms, integration with multiple model providers, and access to streaming API capabilities.

How to add full observability to your LangChain and LlamaIndex agents in under 10 minutes

Ajah has released new observability integrations for LangChain and LlamaIndex, offering features like hallucination detection, cost tracking, and session tracing. This allows developers to quickly add comprehensive monitoring to their AI agents.

Adding hard per-agent spending limits to LangChain and CrewAI agents - DEV Community

A developer has introduced an open-source cost-control tool for autonomous agents built within the LangChain and CrewAI frameworks, enabling the setting of hard per-agent spending limits.

Cloud & platform providers

Changelog - Vercel

Vercel released multiple product updates including expansion of Blob storage limits, native integration of Workflow SDK Nitro v3, new AI model integrations (Kimi K2.7 Code, DeepSeek models via Azure) on AI Gateway, introduction of HarnessAgent abstraction for agent harnesses, drag-and-drop deployment via Vercel Drop, and expanded integrations with Grok Build and Azure.

AI policy, regulation & governance

OpenAI probed on possible user harm before IPO - Taipei Times

Multiple US state attorneys general have issued a subpoena to OpenAI, initiating an investigation into potential user harm caused by its ChatGPT chatbot. This probe comes ahead of OpenAI's anticipated initial public offering (IPO).

Singapore, ILO Renew ASEAN Labour Pact | Migrant Times

Singapore and the International Labour Organization (ILO) have renewed their two-year labour partnership (2026–2028) aimed at assisting ASEAN countries in addressing challenges related to AI, platform work, and the evolving future of work.

AI agents are reshaping identity governance, and attackers are already exploiting the gap - iTnews

AI agents are fundamentally changing identity governance by introducing new attack vectors, such as credential stuffing and deepfake personas, which current security frameworks cannot adequately address. This necessitates an urgent evolution of access controls and audit trail design.

Industry & market moves

Zuckerberg says Meta made 'mistakes' in AI workforce shift | The Business Standard

Mark Zuckerberg acknowledged that Meta made 'mistakes' in its rapid, AI-driven workforce restructuring initiative that involved laying off 10% of staff and reassigning 7,000 employees. He committed to greater stability and redeployment options for affected workers.

Meta reportedly moves to unwind $2B Manus deal after Beijing’s demand

Meta has reportedly begun the process of unwinding its $2 billion acquisition of Manus, following a divestiture order from Beijing on national security grounds. This marks a significant enforcement action of Chinese tech export controls.

2X Acquires Knownwell To Create AI-Powered GTM Services Platform Valued At More Than $400 Million

2X has acquired Knownwell, an AI engineering platform, in a deal valued at over $400 million. The acquisition aims to create a unified, AI-powered go-to-market (GTM) services operating system for B2B enterprises.

Cycclone, Inc. Announces Agreement in Principle for the Acquisition of AI Robotics Company

Cycclone, Inc. has reached an agreement in principle to acquire a California-based AI Robotics company that specializes in autonomous vehicle operations. The acquisition is pending the completion of formal merger and acquisition processes.

AI product & feature launches

Claude's June 15 Billing Change: Should You Worry? | FindSkill.ai — Learn AI for Your Job

Anthropic is changing its billing model for Claude on June 15, separating interactive chat usage (remaining on subscription) from automated agent/SDK usage, which will shift to a metered credit pool charged at API rates. This change primarily affects developers and enterprises integrating Claude via SDKs.

From Noise to Knowledge: How Lysk's AI-Powered Lore Turns Combat Radio into Actionable Data - Second Line of Defense

Lysk's new AI software, Lore, automates the transcription of combat radio communications into actionable written reports. This allows military officers in the field, even without commercial internet, to receive structured intelligence from raw audio.

Snowflake AI Pulse – June 2026 Product Announcements at ...

Snowflake announced several new AI agent and ML features in its June 2026 product release, including CoWork, a personal agent for knowledge workers, and CoCo, a data-native coding agent. These additions expand Snowflake's capabilities for AI-driven productivity and development.

Research with immediate practical relevance

Google Researchers Introduce 'Faithful Uncertainty' to Reduce Hallucinations in Large Language Models - Archynewsy

Google researchers have introduced a new technique called 'Faithful Uncertainty' aimed at reducing hallucinations in large language models (LLMs). This method aligns an LLM's expressed uncertainty with its internal confidence, preserving utility while making the model's outputs more reliable.

Chinese AI Catches Up on Safety-Test Awareness, Neo Research Finds — June 2026

A benchmarking study by Neo Research found that Chinese frontier AI models (DeepSeek V3.2, Moonshot Kimi-K3, Zhipu GLM-5) have rapidly achieved near-US levels of 'evaluation awareness' – the models' internal understanding that they are being safety-tested. This challenges the historical assumption of neutrality in AI safety evaluation frameworks.

Microsoft Research's Mirage gives video generation a persistent spatial memory that doesn't forget what's around the corner

Microsoft Research's new Mirage video world model significantly improves video generation by achieving 10.5 times faster performance and using 55 times less memory compared to similar models. It does this by storing diffusion features directly in latent spatial memory, rather than relying on pixel-based 3D point clouds.

μ₀ World Model Predicts 3D Motion Traces For Efficient Robot Learning

The μ₀ World Model, developed by Furong Huang et al., proposes using 3D motion traces as a physical language for robot learning. This approach achieves performance comparable to action-trained policies but requires approximately 1/100th of the data and no action labels.