🏆 Google's Gemini 3 Pro sweeps 2025 AI rankings 🏆
Tops text, vision, and search categories in LMArena results

LMArena has released its 2025 leaderboard, showing Google's Gemini 3 Pro dominating across multiple categories including text generation, vision tasks, and search capabilities. The results also revealed Google's Veo 3.1 models leading video generation rankings, marking a significant year for Google's AI portfolio. The comprehensive benchmark provides the most complete picture yet of how frontier models stack up against each other in real-world performance.

🤖 OpenAI races to ship audio-first device in 12 months 🤖
OpenAI bets on voice interfaces as current audio models lag behind text-based ChatGPT.

OpenAI has reorganized internal teams to accelerate audio AI development, with plans to launch an audio-first personal device within 12 months. The move aligns with a broader Silicon Valley shift toward voice-first interfaces, as Meta, Google, and Tesla integrate conversational audio into everyday products. Current voice models reportedly lag behind text-based ChatGPT in accuracy and speed—a gap OpenAI aims to close. An upgraded model expected in Q1 2026 will introduce "barge-in" capability, letting users interrupt and redirect the AI mid-response without breaking conversation flow. This signals OpenAI's intent to compete directly in consumer hardware, where pricing and distribution will determine mass adoption.

🇨🇳 DeepSeek publishes super-efficient training method before R2 launch 🇨🇳
Chinese lab advances efficiency research amid hardware export restrictions.

Chinese AI lab DeepSeek has published research on a more efficient training approach using Manifold-Constrained Hyper-Connections (MHC). The framework aims to improve scalability while reducing computational and energy demands—a key competitive advantage given US hardware export restrictions facing Chinese companies. The publication timing suggests DeepSeek is preparing a major model release soon. Industry sources expect the R2 model to launch around February's Spring Festival, potentially challenging Western frontier models on efficiency metrics. This matters for global AI competition, as efficiency gains could help Chinese labs close capability gaps despite chip access limitations.

INTERESTING TO KNOW

🧠 Meta acquires Manus AI 🧠

Meta has acquired Manus AI, a fast-growing “general AI agent” startup. While terms were undisclosed, it is widely reported as a multi‑billion‑dollar deal. This move doubles down on Meta's commitment to agentic AI and recurring revenue models.

Anthropic's Claude Code now writes 100% of its own code

Anthropic has revealed that its agentic coding tool, Claude Code, now generates all of its own code without human input. Boris Cherny, one of the main developers of Claude Code, confirmed that the agent has achieved full self-authorship in the past month.