AI News Roundup: Meta Muse Spark 1.2, OpenAI Updates, and Cloudflare Agents Week

This roundup covers significant updates in artificial intelligence across frontier models, agent infrastructure, open-weight ecosystem tools, and domain-specific benchmarks. Meta introduced Muse Spark 1.2, achieving high performance and low test costs while claiming gold-medal results in several STEM Olympiads using multi-agent orchestration without external tools. OpenAI restructured ChatGPT by unifying instant and reasoning capabilities into GPT-5.6 Sol for paid users, expanding free access to GPT-5.6 Luna, and launching Agent Plugins as an open standard for agent skills and Model Context Protocol configurations.

Cloudflare advanced agent infrastructure during its Agents Week by unveiling Kitesurf, a stateless browser built on Workers designed to reduce overhead for automated agents. Meanwhile, Model Context Protocol integration expanded across various platforms, including a built-in endpoint added by Weaviate. The broader industry debate increasingly centers on whether intelligence lives primarily in foundational models or in inference-time orchestration harnesses like Prime Intellect’s newly released Prime Agent.

Open-weight model availability and local execution also progressed significantly. Preparations advanced for the open release of Qwen 3.8 Max with 2.4 trillion total parameters, while mainline llama.cpp added native support for Qwen3-TTS to enable local voice cloning. In scientific and physical applications, Google DeepMind published WeatherNext 2 to improve tropical cyclone forecasting, and Reka released a large dataset of unscripted household footage to support physical AI development.