Google has officially released the Gemini 3 Deep Think mode for Ultra premium users. It supports multi-path parallel reasoning and significantly improves problem-solving and planning capabilities in mathematics, science, and complex code scenarios. It demonstrates outstanding performance on challenging benchmarks like ARC-AGI-2, featuring multi-round self-iterative analysis, albeit with slower inference speeds and higher resource consumption. It is currently positioned as a paid premium service for high-value scenarios.

Meta plans to cut the budget for its Reality Labs metaverse business by up to 30% in a large-scale reduction, which may involve layoffs, and shift more resources towards AI models and smart glasses. This marks Meta's first comprehensive strategic adjustment since its 2021 rebranding, reflecting the cooling of metaverse implementation and the prioritization of AI investment. Wall Street has welcomed this move.

Anthropic has entered into a multi-year, $200 million partnership with the enterprise data cloud platform Snowflake. Models like Claude Sonnet 4.5 will be embedded as the built-in AI engine for Snowflake Intelligence, providing customers with secure enterprise AI capabilities, including multimodal data analysis of text, images, and AI agent development. This highlights Anthropic's strategy of focusing on the large client market and seeking deep integration of AI with data security.

NeurIPS 2025 announced seven annual awards, focusing on AI model output homogenization (Artificial Bee Colony effect), gated attention mechanisms significantly enhancing multi-model capabilities, ultra-deep (1000+ layer) self-supervised reinforcement learning models achieving performance leaps, and diffusion models preventing training set memorization. Some innovations have already been adopted for industrial use by models like Qwen, laying a theoretical foundation for future LLM output differentiation and continuous evolution.

Evaluations by teams including Anthropic on ten frontier large models using a set of 405 historical smart contract vulnerabilities revealed that AI can automatically replicate approximately 207 vulnerability attacks, theoretically 'stealing' simulated funds totaling $550 million. The AI also discovered unpublished novel vulnerabilities, indicating it has reached near-human capability in exploiting logical flaws, posing urgent challenges for blockchain financial security and AI governance.

Based on an analysis of 100 trillion tokens on the OpenRouter platform, reasoning/coherence models already account for half of practical reasoning tasks. Nearly one-third of the market share is held by open-source large models (driven notably by Chinese models). Community 'role-playing' scenarios far exceed programming/writing needs, showing high reuse stickiness. This indicates that reasoning capability, niche use cases, and model diversification are becoming mainstream market focal points.

Apple announced the collective retirement of its AI head, UI design lead, and legal/policy executives. Key positions are being filled by former Google Gemini and Meta executives. The AI team will be merged into the software division led by Craig Federighi. This releases a strong signal of facing competitive pressure in the AI race, accelerating AI capability transformation, and absorbing external industry resources.

The newly released SUSVIBES security benchmark evaluated multiple mainstream AI programming agents, finding that even when generating functionally correct code, automated code only about 1/6th was safe and compliant. Critical issues like injection vulnerabilities are easily overlooked, and simple prompt engineering struggles to improve security. This warns that security must be designed as a core objective for AI programming agents, not an external patch.

Frontier AI models completed a series of automated offensive and defensive analysis tasks on the security operations dataset Splunk BOTSv3. Among them, GPT-5.1 and Opus 4.5 achieved 63% accuracy. Opus had higher costs but faster speeds, while the Gemini model failed to complete multiple tasks. This indicates clear differentiation in the security automation capabilities and performance in security operations scenarios among different large models.

In a copyright dispute case with The New York Times and several publishers, a US court ruled that OpenAI must publicly release approximately 20 million anonymized ChatGPT chat logs within a time limit. Copyright of AI training data, privacy protection, and the legality of AI-generated content have become the core of the dispute. This is expected to push forward the compliance process for large model data sources and influence industry data governance policies.