The New Bottleneck

Listener notes

• On August 13, OpenAI said its Ultrafast preview runs GPT-5.6 Sol at up to 14 times Standard speed, reaching up to 750 output tokens per second through Cerebras. Access is limited to an initial group of customers, and those figures are company-reported. 1
• On August 14, Anthropic said future Claude models will carry a text watermark that readers cannot see and that adds no hidden characters or extra tokens. The company also says detection becomes weaker on short passages and can show only that Claude was likely involved, not who used it or who authored the text. 2
• In an August 12 arXiv preprint, the PACE-SIMS authors report an AI agent running a blind, randomized 8.1-hour materials study with 35 measurements. They say it made three unscripted corrections, matched all four sealed predictions, and produced transferable calibration results with less than two hours of researcher attention; this is a preprint, not a peer-reviewed finding. 3