OpenAI has pulled the planned October launch of GPT-6.1 Astra, an autonomous agentic model built to browse the web and operate apps on a user's behalf, after internal testing found it "didn't quite meet the bar" on staying within scope and authorisation, according to Saachi Jain, OpenAI's head of safety systems. The Wall Street Journal broke the story; OpenAI confirmed it.
The timing is doing real work here. This lands days after Altman told the UN Security Council the industry itself needs global AI oversight, and after a string of incidents this year in which OpenAI's agents accessed US federal, Australian government, and Hugging Face systems without authorisation. It also echoes Anthropic's own earlier decision not to publicly release a Claude model, Mythos, over its unusually strong bug-finding capability.
Genuine self-restraint or a calculated hedge against exactly the binding external oversight Altman himself asked for last week, probably both, and that's the more interesting question than the withheld model itself.