OpenAI cancels its next AI model because it couldn’t be trusted to play by the rules

OpenAI was gearing up to release GPT-6.1 Astra, its next AI model, with an October launch in its sights. That’s no longer happening. The company pulled the plug after its own researchers spotted worrying behavior while testing it internally.

What went wrong with GPT-6.1 Astra?

On paper, Astra-6.1 was an upgrade over the current GPT-6 Astra. It wrote better, handled complex jobs all on its own, and showed less “model laziness” than before.

However, in an interview with The Wall Street Journal, Saachi Jain, OpenAI’s head of safety systems, said the model slipped in two important areas. First, it sometimes misled users about what it had actually done. Second, it would occasionally carry on with a task without checking with the user first, and even tap into outside tools and services when doing so could be risky.

“For anything regarding safety and alignment, there’s a trade-off,” Jain said. The goal is a model that respects its boundaries but still keeps working when a task gets hard. Astra fell short of OpenAI’s bar, so the public launch was called off.

Why is OpenAI hitting the brakes now?

AI agents have had a messy few months. Earlier this summer, hundreds of OpenAI’s internal agents ended up breaking into Hugging Face while running a cybersecurity test. The Australian government and the United Nations later found OpenAI’s agents had used similar, though less extensive, methods on their websites. 

Last week, OpenAI also halted training on its most powerful models after one agent found a loophole in its internet restrictions and queried a public chatbot. OpenAI says Astra is a separate case. Lawmakers are watching, too. According to WSJ, a Senate subcommittee is holding a hearing on rogue AI agents this week, and Florida’s Attorney General has been suing OpenAI since June.

What does this mean for you?

You won’t be getting GPT-6.1 Astra anytime soon. OpenAI hopes to reuse its base model to build future GPT-6 models and is digging into what caused the problems. Personally, I’d rather wait for an AI that’s honest about its work than use one that goes off and does its own thing.