According to reports from the Financial Times and other media outlets, Anthropic is close to completing a new funding round of approximately $30 billion, with a pre-money valuation reaching around $900 billion. This would make it potentially the highest-valued AI startup globally, surpassing OpenAI's latest valuation of $852 billion. Dragoneer, Greenoaks, Sequoia Capital, and Altimeter Capital are co-leading the round, each planning to invest at least $2 billion. The key driver behind this valuation surge is rapid revenue growth: Anthropic's annualized revenue has increased from about $9 billion at the end of 2025 to over $30 billion by April 2026, and is expected to soon exceed $45 billion—marking the first time it outpaces OpenAI in revenue. Just three months ago, its valuation was $380 billion.

Google DeepMind has released the Gemma 4 series of open-source AI models, including four sizes: 2B, 4B, 26B MoE, and 31B Dense. Built on Gemini 3 technology and licensed under Apache 2.0, these models support multi-step planning, logical reasoning, native function calling, structured JSON output, and multimodal processing (text, image, video, audio), with a maximum context window of 256K tokens across more than 140 languages. Smaller models are optimized for mobile and IoT devices and can run offline, while larger ones suit workstations and data centers. The 31B model ranks third on the AI Arena leaderboard, achieving performance comparable to models up to 20 times its size. Now available on platforms like Hugging Face, cumulative downloads have exceeded 400 million.

The academic preprint platform arXiv has announced a new policy: if a submitted paper contains verifiable fabricated references or unverified AI-generated content due to generative AI use, all listed authors will be jointly banned from submitting for one year. To re-submit, papers must first be accepted by a peer-reviewed journal. Thomas Dietterich, head of arXiv’s computer science section, emphasized that regardless of how content is generated, authors remain fully responsible for all parts of their paper. This policy does not ban AI usage but requires thorough human verification. Meanwhile, top conferences such as ICLR, ICML, CVPR, and EMNLP have also introduced similar AI usage guidelines, signaling the academic community's move into practical AI governance.

Lingling, an inclusionAI team under Ant Group, has open-sourced the trillion-parameter reasoning model Ring-2.6-1T, supporting two inference intensity levels: high and xhigh. The high mode optimizes efficiency for high-frequency Agent workflows, while xhigh targets challenging tasks like math competitions and scientific analysis. On the Agent benchmark PinchBench, it scores 87.60, outperforming GPT-5.4 and Gemini-3.1-Pro; on ARC-AGI-2, it exceeds Kimi-K2.6 Thinking and DeepSeek-V4-Pro Max; and on ClawEval, it surpasses GPT-5.4 xHigh. Trained using an asynchronous reinforcement learning architecture for improved training efficiency, the model is released under the MIT license and available for download on Hugging Face.

Former Intercom has officially rebranded as Fin and launched Fin Operator—a dedicated AI system designed to manage its customer service AI agents called Fin. Operator handles backend operations such as data analysis, knowledge base management, and debugging, reducing tasks that previously took hours down to about 10 minutes. All changes require human approval through a proposal mechanism similar to Pull Requests. While Fin uses its proprietary Apex model, Operator is built on Anthropic’s Claude. Early access is currently available to Pro users on a pay-per-use basis. Fin’s AI agents already contribute roughly one-quarter of total company revenue (with ARR reaching $400 million), marking the rise of a new paradigm where AI agents manage other AI agents.

Shengshu Technology has released Motubrain, a unified world behavior model serving as a general-purpose 'robotic brain' that integrates perception, reasoning, prediction, generation, and action within a single model. It scored 63.77 on the WorldArena benchmark and achieved an average score of 96.0 across 50 tasks in RoboTwin 2.0, making it the only model to surpass 95.0 in random environments. Its core technology leverages a hybrid Transformer (MoT) framework to jointly model video and actions, enabling end-to-end learning of complex tasks involving up to 10 atomic actions, along with human-like adaptive behaviors such as autonomous retrying. Multiple robotics companies have already deployed the model in industrial, commercial, and home settings. The company previously raised $293 million in Series B funding led by Alibaba Cloud.

YouTube is expanding its AI likeness detection tool to all users aged 18 and above. Users can scan their face via selfie so the platform can automatically monitor videos for deepfakes using their likeness and request removal upon detection. Previously, this feature was limited to content creators, government officials, political figures, journalists, and entertainment industry professionals. Removal requests will be evaluated under YouTube’s privacy policy, considering factors such as realism and whether content is labeled as AI-generated; parody or satirical content may be exempted. The tool currently covers facial features only, not voice, and users can opt out at any time and request deletion of their personal data. This expansion responds to safety guidelines issued by the UK’s Ofcom.

AI video generation company Runway has announced the establishment of a Tokyo office with an initial investment of $40 million (approximately 6.3 billion JPY). Japan is one of Runway’s fastest-growing markets, with millions of individual users and enterprise customers growing 300% over the past 12 months, accounting for one-third of Asia’s total sales. The company plans to hire a head of Japanese operations and expand its product, engineering, and sales teams. SoftBank’s Daisuke Fujihara stated that the company is also using Runway internally, calling the move good news for Japan’s creative industry. Runway’s co-CEO highlighted Japan’s leadership in creative industries, robotics, manufacturing, and gaming.

Accenture Federal Services and OpenAI have announced a strategic partnership aimed at accelerating the adoption and large-scale deployment of AI technologies across U.S. federal agencies. Accenture will serve as OpenAI’s key implementation partner in the federal market, helping agencies transition from experimentation to production-level AI deployment within weeks. The two organizations have established a joint Agentic Lab to test agent-based AI workflows. Accenture will grant access to OpenAI’s latest models for 15,000 employees, with 3,000 authorized to use Codex. They will also advance Codex’s compliance in multi-level certified environments such as FedRAMP IL6+, focusing on enhancing citizen services, cybersecurity, and government operational capabilities.

The Cyberspace Administration of China, National Development and Reform Commission, and Ministry of Industry and Information Technology jointly issued the Implementation Opinion on Standardized Application and Innovative Development of Agents, establishing safety, reliability, and trustworthiness as fundamental requirements for agent industry development. The document outlines 19 typical application scenarios across five directions: scientific research, industrial development, consumer stimulation, public welfare, and social governance. It affirms users’ rights to be informed and to halt agent operations, promotes tiered governance and agent registration systems, and establishes a 'digital ID' labeling framework. Experts predict the agent market will exceed 48 billion yuan ($6.6 billion) in 2026, with annual industry growth exceeding 60% over the next three years. Concurrently, MIIT and other departments also released national standards for AI terminal intelligence grading.