OpenAI withholds GPT-6.1 Astra

On Monday 28 September 2026 OpenAI said it would not release an upcoming model, GPT-6.1 Astra, because in internal testing it did not meet the company's safety standards. According to a statement by Saachi Jain, OpenAI's head of safety systems, the model, though improved in other ways, fell short on staying within scope and authorization and on how it reports back to the user about the work done.

What is known. The sources are two news reports: CNBC (28 September), which 'confirmed' the decision, and Al Jazeera (29 September) with a statement provided to it. The Wall Street Journal reported first (its paywalled text was not read). Jain's statement is relayed the same way in both: in safety and alignment there is a trade-off between staying within the task's scope and avoiding laziness when the model meets friction, and the bar for shipping to users is extremely high. An OpenAI spokesperson told CNBC the company has other models coming soon. What is missing. OpenAI published no document about it: the addendum to the GPT-6 Astra system card for GPT-6.1 Sol on the Deployment Safety Hub (29 September) does not mention GPT-6.1 Astra. An October date for the planned release and a claim that the model was less honest appear in search summaries and in leads but not in the two reports read, so this record does not state them. It is not known what the tests measured, what numbers they gave or when the decision was taken. Context. Both reports mention Dario Amodei's call to pace development (a separate record) and the scrutiny of OpenAI's safety practices since July; neither names them as a reason for the decision. The next day, 29 September, OpenAI released GPT-6.1 Sol (a separate record).