Nvidia’s robotics chief Deepu Talla publicly praised OpenAI’s GPT-6 Astra as a win for robotics, saying a model built to understand and use language could help machines move and work in the real world. The comments came at an awkward moment for Nvidia: Jensen Huang recently stepped down as CEO amid scrutiny of his conduct, and weakness in tech stocks has been tied to uncertainty around his exit. Nvidia has also faced attention over Megaspeed, a Singapore-based company that bought its chips. The U.S. Departments of Commerce and Justice issued a rare joint statement clarifying their investigation. According to the statement, the Commerce Department opened an inquiry in 2024 into possible export control violations, asked Nvidia in May 2025 for records on direct or indirect sales to Megaspeed, and in June 2025 Nvidia placed Megaspeed on its “no sell” list. That same month, Bureau of Industry and Security agents worked with Singaporean law enforcement on an on-site check and interviewed Megaspeed CEO Alice Huang. The Justice Department then took the lead in a criminal investigation, and officials said 17 people across six cases have been charged under the Trump administration over alleged smuggling of billions of dollars’ worth of chips. The statement does not accuse Nvidia of wrongdoing.
Independent evaluations, however, show Astra is not yet ready for real-world robotics deployment. In the AA-Robotics v0.1 research preview from Artificial Analysis, GPT-6 Astra scored 69 and ranked first among five top AI models, but the models were about 100 times slower than a person. A human needed about 20 seconds per task, while Astra needed about half an hour, and Astra was the most expensive at $16.36 per task. RoboCurve found Astra could put a block in a bowl in 19 out of 20 tries, a 95% success rate at $0.94 per try, but on a harder precision task of pressing a puzzle piece into a tight round slot, Astra succeeded only 2 times in 20, or 10%, the same as Anthropic’s Claude Fable 5.1. Nvidia’s Seattle Robotics Lab “Astranomy” study scored 74.8% on RoboLab-120, 23.8 points higher than the previous best method, but each decision took 18.56 seconds and one successful run used 30.9 million tokens at about $36.40. The study itself said delay and cost remain major obstacles.
OpenAI followed with the release of GPT-6.2 Astra and Sol, promising stronger professional work at lower task costs. Astra targets demanding projects such as financial spreadsheets, legal research and interactive 3D design, while Sol handles everyday work at a lower price. Astra costs $8 per million input tokens and $40 per million output tokens, down 20% from the original model; Sol retains $2 and $10 rates. Independent testing by Artificial Analysis placed Astra at 62 on its Intelligence Index, just behind Anthropic’s Claude Fable 5.2 at 65, while Sol scored 57 and matched the original Astra’s performance at roughly 15% of its cost per evaluated task. Alibaba’s Qwen 4 Max entered the conversation at $1.50 per million input tokens and $4.50 per million output tokens. OpenAI said Astra is its first externally released model to reach the High threshold for AI self-improvement and that both new models meet its Critical cybersecurity threshold, with a new Control Room system to halt suspicious model activity. Nvidia has long promised a robotics breakthrough: at CES in January, then-CEO Jensen Huang said the “ChatGPT moment” for physical AI was “nearly here.”