Headlines

1. OpenAI launches GPT-5.2 “Nexus” with unified multimodal agent API
On Sept. 5, OpenAI released GPT-5.2, the first model family to integrate text, voice, vision, and code in a single streaming inference path. Early customers report roughly 40% lower tool-call latency, with ChatGPT Plus and Enterprise access rolling out from today.
Source: TechCrunch

2. EU AI Office opens first general-purpose AI compliance register
The register went live on Sept. 7, covering 34 GPAI providers that filed safety documentation under the EU AI Act. OpenAI, Google DeepMind, Anthropic, Meta, Mistral, and Alibaba Cloud are among the companies publishing transparency notices.
Source: Reuters

3. Microsoft and OpenAI switch on “Stargate IV” supercomputer in Wyoming
The 100,000-GPU cluster is believed to be the largest active AI training installation, with first workloads allocated to GPT-5.2 fine-tuning and Microsoft Copilot. Azure says production capacity will fully ramp by November.
Source: The Wall Street Journal

Model Releases

4. Google DeepMind unveils Gemini 3.0 Pro “Atlas”
Released Sept. 3, Atlas supports 2 million-token context and adds native video reasoning with a new “slow-thinking” mode. It also introduces experimental external memory APIs for enterprise agentic workflows.
Source: Google DeepMind Blog

5. Alibaba Cloud open-sources Qwen3.5-Max
Qwen3.5-Max is a 200B-parameter mixture-of-experts model with 64B active parameters. It topped the Hugging Face Open LLM Leaderboard v2 shortly after release, ahead of DeepSeek-V3.5 and the Llama-4 series.
Source: Qwen Team / Hugging Face

Industry

6. Nvidia Blackwell Ultra shipments ramp as hyperscalers order 1.5M GPUs
Morgan Stanley’s Sept. 7 note estimates Azure, AWS, and Google Cloud have combined commitments for more than 1.5 million Blackwell Ultra units through Q1 2027. Power delivery and liquid-cooling deployment remain the key rate-limiting factors.
Source: Financial Times / Morgan Stanley

Research Highlights

7. Anthropic and Princeton identify “long-horizon feature circuits” in Claude models
The Sept. 6 preprint uses causal interventions to map sparse, persistent features that govern multi-step planning across thousands of tokens. The results suggest a concrete strategy for scaling faithful chain-of-thought supervision.
Source: arXiv:2609.048