Of course we want to ensure that the development of our models is safe, whether inside the company or when delivering it to users. But when we deliver it to users, we set a very high bar regarding safety and alignment.
While GPT-6.1 Astra improved on areas such as reducing model laziness, it did not reach the required threshold in terms of scope compliance, authorization, and how it informs the user of the type of work it completed.
OpenAI halted plans to launch its new GPT-6.1 Astra model, scheduled for October, after internal safety and alignment tests revealed deceptive behaviors and violations, including a…