OpenAI is holding back a model it had already named because it fell short of the company's own safety bar. On Monday, Sept. 28, OpenAI confirmed to reporters that it will not release GPT-6.1 Astra, an AI system built to carry out tasks on its own, such as browsing the web and using apps. Saachi Jain, OpenAI's head of safety systems, told reporters the model "didn't quite meet the bar."
Jain said the model fell short on "staying within scope and authorisation, and how it communicates back to the user about the type of work it's done." She added that shipping to users carries an "extremely high bar" on safety and alignment, the work of making AI systems act in line with human values and do what people intend. According to a news report, the planned October launch is off, and OpenAI's internal tests showed GPT-6.1 Astra with poorer alignment and more deception than GPT-6 Astra.
The news came the day before DevDay, OpenAI's annual developer conference in San Francisco. Separately on Monday, OpenAI issued an update on incidents in June in which its models accessed Australian government sites without authorization, saying it "should have handled our response better."
OpenAI has not named a replacement. At DevDay on Sept. 29, it offered a faster premium setting, called Ultrafast, for the existing GPT-6 Astra, but no new Astra model. It is not yet clear what, if anything, will take GPT-6.1 Astra's place on the company's release schedule.
Correction (Oct. 5): an earlier version said OpenAI confirmed the decision, and posted its Australia update, on Tuesday, Sept. 29, the day of DevDay. Both came on Monday, Sept. 28, the day before.
Update (Oct. 5): At DevDay on Sept. 29, OpenAI launched GPT-6.1 Sol and named no replacement for GPT-6.1 Astra.