AI News Roundup: OpenAI and Anthropic Agent Updates, AT&T Enterprise Routing, and Rubin Infrastructure

This report summarizes artificial intelligence updates, product releases, model benchmarks, and open-source developments from mid-August 2026. Major labs expanded their agent capabilities and desktop integrations. OpenAI introduced Apple Messages and collaborative editing features for Mac users, while Anthropic made its computer use, browser tool, and Skills and Files APIs generally available. Meta advanced its multimodal Muse Spark 1.2 model, and Google highlighted the cost-adjusted reasoning performance of Gemini 3.7 Flash across multiple benchmarks.

In enterprise and model economics, AT&T emerged as a key case study for hybrid routing, directing 40 percent of its employee AI usage to open models while cutting coding costs significantly. Meanwhile, open-weight adoption expanded as models like Qwen3.8-27B and the Ornith-1.5 family gained traction for coding and agentic workflows. Developers also implemented inference speedups, such as DFlash2 speculative decoding, to achieve high token throughput on consumer-grade hardware.

Hardware and systems engineering saw notable milestones, including OpenAI installing its first NVIDIA Vera Rubin racks for frontier pretraining. Cerebras launched the CS-4 inference system to scale performance on a single wafer, and researchers explored new methods for persistent agent memory and guarded harness evolution to improve reliability.