Anthropic acquired Stainless, a leader in SDK generation and MCP server tooling, to enhance its agents' ability to connect with external systems. The acquisition strengthens Anthropic's shift from response models to acting agents, enabling Claude to access more data and tools.

OpenAI partnered with Dell to bring AI coding models to hybrid and on-premise enterprise environments. The alliance enables companies to securely deploy AI agents within their own systems, without relying on cloud infrastructure.

KPMG, the professional services firm with 276,000 employees, announced a strategic alliance with Anthropic to integrate Claude into Digital Gateway, its central work platform. All KPMG employees will have access to Claude for audit, tax, and legal tasks.

Mistral AI completed the acquisition of Emmi AI to strengthen its AI platform. The acquisition reinforces Mistral's ability to offer a complete and competitive stack in the European market.

Ocean, an agentic email security platform, raised $28 million from Lightspeed Venture Partners. The startup addresses the growing problem of AI-driven phishing.

At Google I/O 2026, the company announced the shift toward an agentic agent era with Gemini. This includes multiple products and tools designed to automate tasks and enhance developer productivity.

Google revolutionized its iconic search box to accept text, images, PDFs, videos, and tabs. The redesign integrates AI Overviews and merges traditional search with AI-driven experiences in a unified flow.

SandboxAQ integrated its drug discovery models with Claude to make them accessible to users without computing expertise. The company believes access, not just model quality, is the main barrier in AI applied to sciences.

Google unveiled multiple innovations at I/O 2026 including new Gemini models, search redesign, and smart glasses. The event solidified the company's strategy toward a more AI-integrated experience.

Google presented its vision for the future of AI at the I/O 2026 conference. The event highlighted how the company is integrating AI across its products to make it more useful for users.

Google expanded Gmail with conversational voice search using Gemini. Users can now ask natural language questions to find buried information in their inbox.

Google launched AI agents that monitor topics and proactively alert users to updates. These agents go beyond traditional searches to deliver automatic updates on topics of interest.

Google and Volvo integrated Gemini into the EX60 SUV to interpret parking signs. The model can now access external vehicle cameras to explain its surroundings to drivers.

NVIDIA began delivering Vera, its first CPU specifically designed for AI agents, to leading labs including Anthropic, OpenAI, and SpaceX. Vera promises to run agent inference more efficiently than traditional CPUs.

The Open Agent Leaderboard was launched, an open benchmark for evaluating multi-agent system performance. The tool allows developers to compare agent implementations and share best practices.

NVIDIA presented Vera Rubin NVL72 results showing agentic inference at one-tenth the cost of traditional tokens. The chip runs agent sandboxes 50% faster and accelerates enterprise queries up to 3x, with 5,000 companies already using Dell AI Factories.

Hugging Face launched Ettin, a new family of reranking models that improve precision in search and information retrieval. These models are useful for RAG applications and production search systems.

NVIDIA and Google Cloud expanded their joint program for over 100,000 developers. The initiative provides curated learning paths and hands-on labs to build with NVIDIA platform on Google Cloud.

Hugging Face published guides for fine-tuning NVIDIA Cosmos Predict 2.5 with LoRA/DoRA for robot video generation. This advancement makes it easier for researchers to adapt video models for specific robotic tasks.

Google redesigned its AI creation software Flow with a new video model and avatar generator. The tool enables easy selfie video creation, though it raises concerns about deepfakes.

PaddleOCR released version 3.5 with Transformers backend support, improving OCR and document parsing tasks. The update enables developers to use Transformers models for more accurate optical recognition.

Google unveiled Gemini 3.5, its new frontier model with enhanced action capabilities. The release marks a significant step forward in the evolution of the company's multimodal models.

Hugging Face and Allen AI released OlmoEarth v1.1, a more efficient family of models. The update improves performance while maintaining lower computational requirements, benefiting developers with limited resources.

ByteDance released Lance, an open-source multimodal model with only 3 billion parameters handling image generation, video, and editing. Trained from scratch with limited budget, Lance proves extreme efficiency is possible in multimodal AI.

OpenAI deployed Content Credentials and verification tools to combat misinformation. The initiative aims to help users identify and trust AI-generated content safely.

The Take It Down Act came into full force, requiring social networks to quickly remove sexual deepfakes and nonconsensual intimate imagery. Experts warn the regulation could have limited impact on victims and potentially facilitate online censorship.

Former OpenAI employees founded a watchdog group and warned that xAI has a poor safety record that could complicate SpaceX's IPO. Ex-colleagues urge investors to demand greater transparency on safety practices before the public offering.

Google presented its AI vision at I/O 2026, highlighting tools like Gemini Spark and Daily Brief that require personal data access. The approach raises privacy questions in the agentic era.

An arXiv paper analyzes how multi-agent systems scale with 15 frontier LLMs and 1,141 real skills, identifying coupled routing and execution laws. Findings show routing accuracy decays logarithmically and propose optimizations to improve agent performance.

Researchers presented SDOF, a framework treating multi-agent execution as constrained state machines. In tests with 6,000+ enterprises, SDOF achieved 86.5% task completion while blocking malicious operations, outperforming GPT-4o in specialized routing.