Here are ten AI headlines for October 28, spanning faster models, re-specced chips and agents looking for real entry points across the industry.

__FIG__

1. OpenAI ships GPT-6.1 Terra. Per Sina and MyDrivers, the model can operate a computer on its own; multiple reports cite roughly 3x flagship speed at sharply lower cost.

2. NVIDIA trims Vera Rubin specs. Expreview reports HBM4 downclocking and reduced NVLink bandwidth on the upcoming platform, a sign that even next-generation racks must balance cost against memory supply.

3. RTX 5090 shortage chatter. Per Phoenix Tech and Sina, partners reportedly cannot secure core chips, while Huaqiangbei prices climb past RMB 50,000 per high-end card.

4. Hidden GPU power unlocked. 36Kr reports developers bypassing official drivers to double RTX 4090 throughput.

5. Agents hit platform walls. Per Sohu and DoNews, Muse and Doubao agents face access restrictions from banking and social apps, turning platform openness into the newest friction point for agent deployment.

6. Anthropic report: agents work 6x longer. Sin Chew via China.com cites a six-fold effective-hours figure versus humans on some tasks, driving fresh demand for enterprise AI safety tooling.

7. Wall Street joins the AI talent war. Sina reports quant firms recruiting on campus against OpenAI and other labs.

8. China's intelligent computing ranks second globally. Sohu cites over 15.56 million racks in use with fresh project milestones.

9. OpenAI publishes math and physics results. Sohu's global markets column notes new basic-science outputs from the lab, extending the AI-for-science momentum into a second year.

10. Ningxia rolls out power-for-compute measures. People's Daily Online cites full-chain support covering green power, grid connection and pricing.

Faster models, re-specced silicon and agents searching for entry points all point the same way: capability supply is outrunning application governance, and the gap between what systems can do and what platforms allow is now the industry’s main scheduling problem. (Compiled from public reports)