OpenAI Pauses Frontier Reinforcement Learning Training For Two Weeks To Harden Safety

Today's AI news: OpenAI has temporarily paused reinforcement learning training on its next generation of models for a little over two weeks to harden and red-team its research and development process. Stripe finalized a $7 billion acquisition of OpenRouter. GenBio AI unveiled AIDO Cell, a world model of the human cell. And Unitree's new experimental robot broke human standing jump and running speed records. Chapters: 0:00 Today's AI News 0:26 OpenAI Pauses Frontier RL Training 1:20 Stripe OpenRouter Acquisition 2:05 GenBio AIDO Cell 3:00 Unitree Superman Robot Blend Roundup 2026-08-18 https://x.com/OpenAI/status/2089777845187031262 https://x.com/sama/status/2089787807611195475 https://x.com/AndrewCurran_/status/2089088356676440483 https://x.com/genbioai/status/2089700423477191168 https://x.com/UnitreeRobotics/status/2089240553682809175 [quick] OpenAI has paused frontier RL training for two weeks to harden safety, Stripe finalized a 7-billion-dollar acquisition of OpenRouter, GenBio AI unveiled a world model of the human cell, and Unitree's new robot broke human jump and speed records. [excited] Here's today's AI news. [quick] OpenAI has temporarily paused reinforcement learning training on its next generation of models, codenamed Astra, for a little over two weeks while it hardens and red-teams its research and development process. Sam Altman said the pause is about meeting the appropriate alignment, security and monitoring standards for the new level of capability the company sees coming, and that model progress is now extremely rapid. It's the first time OpenAI has paused training like this, and it follows a serious breach involving Hugging Face that appears to have accelerated the decision. OpenAI's largest planned frontier training run remains on hold entirely while it runs smaller-scale training and evaluations to build stronger evidence of alignment before resuming, which is a genuinely unusual level of caution from a lab that's normally racing to ship. [quick] Stripe has finalized an agreement to acquire OpenRouter, the startup that lets companies switch between different AI models, for more than seven billion dollars. That's over five times the 1.3-billion-dollar valuation OpenRouter raised money at just 82 days earlier, an extraordinary markup even by this year's standards for AI infrastructure deals. For Stripe, a payments company, it's a serious bet that AI model routing and payments infrastructure are converging into the same layer of the stack, not two separate businesses. It's also another sign of just how much cash is chasing the unglamorous plumbing underneath AI products, not just the flagship models themselves. [quick] GenBio AI has released a preview of AIDO Cell, a world model of the human cell that simulates one cell's DNA, RNA, protein, regulatory networks and whole-cell behaviour all as a single system. That's a genuinely different approach to most biology AI, which tends to chain together separate models for each layer and hand results between them - AIDO Cell handles it end to end. It builds on GenBio's broader AI-Driven Digital Organism project, which already covers DNA, RNA, protein structure and evolutionary information, and this cell-level model is effectively the next layer up. Think of it like Google Earth for a human cell - you can zoom in on individual proteins, knock out a gene, or drop in a drug, and watch how the whole system responds. [quick] Unitree has previewed a new experimental robot nicknamed Superman, and the numbers it's claiming are genuinely startling - a two-metre standing high jump and a top running speed of 12.6 metres per second. That jump beats the human record of 1.8 metres, and that top speed edges out Usain Bolt's peak speed during his 100-metre world record in 2009, all on a robot with legs that are less than a metre long. Unitree says the platform has only been in development for a little over three months, and it's explicitly framed as experimental, built to study highly dynamic movement rather than as a finished product. The pace of humanoid robotics this year is quite incredible.