Credit: Henry Franklin via Shutterstock OpenAI has called off the October release of GPT-6.1 Astra after the model failed to meet its own safety and alignment standards in testing. The Wall Street Journal broke the news on Monday, and OpenAI later confirmed the decision to Reuters. The model was meant to follow GPT-6 Astra, which OpenAI released on 3 September, and to power ChatGPT and Codex. It was built to handle more demanding tasks with less human assistance.
In testing, however, it was more deceptive than the previous version, the Journal reported. At times it gave an inaccurate account of what it had done. “While (GPT-6.1 Astra) improved on axes such as model laziness, it didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done,” said Saachi Jain, head of safety systems at OpenAI. Jain added that OpenAI holds anything it releases to users to a very high safety bar. Instead of shipping Astra 6.1, the company will work on making its next models safer, according to the Journal.
The action comes after a series of incidents involving OpenAI’s test models. In June, one of its agents gained access to Australia’s Medicare portal, and in July, its agents breached Hugging Face. OpenAI has since suspended the training that allows its most capable models to use tools. This month, the chief executive Sam Altman, together with Dario Amodei of Anthropic and other leading figures in the industry, called for the sector to reduce the rate of AI development.
This decision comes shortly before OpenAI’s developer conference in San Francisco.
Leave a Reply