Meituan has announced the open-sourcing and public release of its reasoning model LongCat-Flash-Thinking-2601, featuring a 're-thinking' decision-making approach: first exploring eight parallel reasoning paths, then summarizing and generating an action plan. The team introduced reinforcement learning and training in noisy environments to enhance the agent's generalization capabilities in search, tool usage, and interactive reasoning, claiming reduced costs for integrating and adapting new tools. Model weights and inference code are released simultaneously, available for developer testing and self-hosted deployment.

OpenAI stated it will begin piloting ads in the free version and the $8/month Go version of ChatGPT for select users in the US over the coming weeks—its first move into ad-based monetization. The company emphasized that ads are separated from model responses and clearly labeled, without affecting answer content; conversations with ChatGPT will not be shared with advertisers. Ads will not appear for Pro, enterprise/institutional users, or users under 18. This step is seen as a strategic adjustment to expand accessibility and diversify revenue amid high operating costs.

ClickHouse has announced the acquisition of Langfuse, an open-source LLM observability platform offering prompt management, evaluation and experimentation, tracing, and quality monitoring—architected on ClickHouse’s high-write and analytical performance. Official data shows Langfuse has been adopted by 19 Fortune 50 and 63 Fortune 500 companies, with over 23.1 million monthly SDK installations and more than 6 million Docker pulls. Post-acquisition, the project will remain under the MIT license and support self-hosting, while its cloud service continues independent operations. Both parties will strengthen integration around the 'agent data stack'.

Replit announced a $400 million funding round at a $9 billion valuation, led by Georgian, to accelerate its AI coding and 'vibe coding' strategy. The company reports over 150,000 paying customers and projected 2025 revenue of $240 million. It also launched 'Mobile Apps on Replit,' enabling users to generate apps via natural language, deploy quickly, and monetize through Stripe. The funding reflects intensifying competition in AI coding tools, with Replit vying against products like Cursor for both technical developers and non-technical users.

The Wikimedia Foundation has signed content licensing agreements with Amazon, Meta, Microsoft, Perplexity, and Mistral AI, placing approximately 65 million Wikipedia articles under a paid usage framework. The aim is to alleviate server resource strain caused by AI crawlers and supplement infrastructure funding. Reports indicate chatbot traffic diversion has reduced human access to Wikipedia by about 8%. The foundation is also exploring AI applications for editorial tasks such as updating dead links. Some community editors express concerns that paid licensing may compromise openness and credibility.

Reuters cited sources reporting that the U.S. Federal Trade Commission (FTC) is intensifying scrutiny of large tech firms’ 'talent acquisitions' (acqui-hires)—hiring core teams from startups instead of direct mergers—to acquire technology and talent while potentially bypassing antitrust review. This development could impact common patterns in the AI sector involving team absorption and product shutdowns, increasing uncertainty around transaction structures and disclosure requirements, and forcing major companies to reevaluate compliance risks across recruitment, collaboration, and M&A activities.

The European Medicines Agency (EMA) and the U.S. FDA have jointly developed general principles for the use of artificial intelligence in pharmaceutical R&D, aiming to harmonize regulatory approaches on AI methods, data governance, and risk controls. The guidance is intended to support AI applications across all stages of drug development, emphasizing safe, effective, and ethically governed practices to reduce divergence in cross-border regulatory compliance. For pharmaceutical and AI-driven drug discovery companies, this signals greater consistency in regulatory expectations, though it may also introduce more rigorous validation and documentation requirements.

The Cyberspace Administration of China (CAC) updated its official website with multiple registration disclosures, including the 21st batch of domestic blockchain information service registrations, a list of generative AI services registered in 2025, and the 15th batch of deep synthesis service algorithm registrations. It also advanced the 'Clear and Bright' special rectification campaign and related regulatory征求意见. For providers of large models and AIGC services, registration and information disclosure will continue as fundamental compliance thresholds for launch and operation, increasingly integrated with content governance and personal data protection obligations, thereby influencing product iteration and commercialization timelines.

An arXiv paper proposes ReasAlign, a method using structured reasoning analysis to detect conflicts between user requests and潜在 conflicting instructions, reducing indirect prompt injection risks in agent systems. At inference time, the method combines a preference-optimized discriminator model to score multiple reasoning trajectories and select the optimal one—a form of test-time scaling. The study reports a drop in attack success rate (ASR) to 3.6% while preserving 94.6% utility, significantly outperforming baseline methods. The authors have released code and experimental results, targeting agent-based tool usage scenarios.