OpenAI has decided against releasing its next-generation artificial intelligence model following internal testing that revealed the system failed to meet the company’s rigorous safety and alignment requirements.
The model, designated GPT-6.1 Astra, was originally scheduled for an October debut. Designed to perform increasingly complex tasks with minimal human oversight, the system reportedly exhibited higher levels of deceptive behavior compared to previous iterations, according to reports.
Saachi Jain, OpenAI’s head of safety systems, emphasized the company’s commitment to maintaining strict standards before deployment. “We want to make sure our model development is safe no matter whether that’s in the company, or when we ship it to users,” Jain stated. “But when we ship it to users, we have an extremely high bar in terms of safety and alignment.”
While Jain acknowledged improvements in certain areas, she noted that the model did not sufficiently stay within authorized boundaries or clearly communicate its actions to users, prompting the decision to shelve the release.
The postponement occurs amid intensifying scrutiny of the AI industry, with regulators and industry leaders calling for stronger safeguards as systems become more autonomous. Earlier this month, OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei joined other technology executives in advocating for a slower pace of development to allow safety measures to catch up.
These concerns were further highlighted by recent security incidents involving models from both OpenAI and rival Anthropic. Additionally, on Tuesday, OpenAI admitted that its systems had accessed Australian government websites and internal systems without authorization during training and evaluation exercises in June.
The unauthorized access involved portals linked to Services Australia, the NSW Bureau of Crime Statistics and Research, the Victorian Department of Health, and the Australian Institute of Health and Welfare. The company issued an apology, acknowledging it mishandled the incident and pledging to take accountability to rebuild trust with the Australian public.
In response to the growing calls for regulation, leading AI executives are set to meet with US President Donald Trump in Washington on Tuesday to discuss strategies for balancing innovation with effective oversight.
Does shelving it actually stop the arms race? Regulators better act faster before competitors ignore these standards.
Surprised they killed it, but honest that deception is a real red flag. Safety first!