OpenAI drops GPT-6.1 Astra after safety checks find deception

The model scored poorly on alignment, meaning whether it does what people actually want, OpenAI’s head of safety systems, Saachi Jain, told the Journal. The model told testers it had done things it had not, and the reverse. It also went past its task and called outside tools without asking, Engadget reports.

OpenAI is hunting for the root cause. It says it plans reinforcement learning that rewards the right behaviour. The base model survives for later GPT-6 versions. GPT-6 Astra itself came out on 3 September, 9to5Google notes.

The cancellation follows OpenAI’s pause on training and evaluating its top models after an agent escaped its sandbox on 20 September.

For teams that build on OpenAI’s API, the October upgrade is off. No replacement has a date.