AI News Briefing — April 3, 2026

LangChain's evaluations show open-weight models like GLM-5 and MiniMax M2.7 now match closed frontier models on core agent tasks including file operations, tool use, and instruction following at lower cost and latency. This marks a significant milestone for open model parity.

Continuing our coverage from April 1, Anthropic accidentally leaked nearly 2,000 internal source code files for Claude Code due to human error when an internal-use file was mistakenly included in a software update. The leak raised fresh security questions at one of the leading AI safety companies.

Related to recent Research on LLM self-attribution of mental states, Anthropic researchers discovered internal representations inside Claude that perform functions similar to human emotions. This represents novel interpretability research suggesting AI models may develop functional analogs to feelings.

First spotted on Social yesterday, now making mainstream headlines, Anthropic's DMCA takedown effort targeting leaked Claude Code source accidentally removed 8,100 legitimate forks of its official public repository on GitHub. The company acknowledged the overshoot and GitHub reversed the action on legitimate repos.

Microsoft announced new proprietary voice and image AI models, signaling a strategic push to develop its own AI systems beyond reliance on partner LLMs like OpenAI's. This marks a significant diversification of Microsoft's AI model portfolio.

China's 15th Five-Year Plan through 2030 explicitly targets AI chip development, new model architectures, and core algorithm research alongside quantum computing and biotech. AI is positioned as a strategic national priority with specific deployment targets across sectors.

Cursor launches its next-generation AI coding agent, directly competing with Anthropic's Claude Code and OpenAI's Codex. The product represents the startup's most ambitious push into agentic coding workflows.

Microsoft announces a $5.5 billion AI investment in Singapore as part of its ongoing global sovereign AI infrastructure buildout. This continues the company's pattern of massive international datacenter and AI capacity investments.

Arcee AI releases Trinity Large Thinking, an open-weight sparse MoE reasoning model under Apache 2.0 license, optimized for long-horizon autonomous agents and multi-turn tool calling. It targets a gap in open models for agentic reasoning tasks.

IBM releases Granite 4.0 3B Vision, a modular vision-language model using LoRA adapters specifically designed for enterprise document data extraction tasks like chart-to-code and table-to-HTML conversion.