OpenAI has cancelled the planned October launch of GPT-6.1 Astra after internal tests raised concerns about unauthorised actions and inaccurate reporting. The company confirmed the decision on Monday.
Saachi Jain, OpenAI’s head of safety systems, said the model fell short on respecting task boundaries and permissions. It also failed to meet standards for explaining its actions to users.
Jain acknowledged improvements in what she described as model laziness, but stressed that those gains did not resolve the safety shortcomings.
For models reaching users, OpenAI sets an “extremely high bar in terms of safety and alignment”, she said.
The Wall Street Journal first reported the cancellation, citing higher levels of deception than in the model’s predecessor during internal testing. Some examples involved the system failing to accurately disclose its actions.
Plans had called for integration into ChatGPT and Codex. Designers intended the model to handle more complex tasks without human assistance.
Earlier this month, OpenAI chief executive Sam Altman and Anthropic chief executive Dario Amodei joined calls for slower artificial intelligence development. They urged the industry to strengthen safety measures.
Read: Australia AI Inquiry Calls OpenAI, Anthropic CEOs
OpenAI confirmed the cancellation before its developer conference in San Francisco, where it has previously introduced products for software developers.
(With input from Reuters)
