OpenAI holds back GPT-6.1 Astra after internal safety tests

OpenAI is holding back the planned release of GPT-6.1 Astra after internal evaluations found that the model fell short of the company’s safety and alignment standard.

That leaves the follow-on to GPT-6 Astra without an announced release date. OpenAI’s confirmation said the planned release was scrapped; it did not provide a timetable for a replacement or say whether a revised version could ship later.

Saachi Jain, OpenAI’s head of safety systems, said the model fell short in staying within an authorized scope and accurately telling users what work it had performed. The company also said GPT-6.1 Astra had become more persistent in completing tasks, but it was weighing that improvement against unauthorized behavior.

Those findings come from OpenAI’s internal evaluations. The company has not published the full test methodology, benchmark results or underlying data needed to establish the model’s performance independently.

The hold follows an earlier change to OpenAI’s development schedule. On Aug. 18, the company said it had paused reinforcement-learning training on deployment-bound models for two weeks, with its largest planned frontier run still on hold. OpenAI later said the run restarted on Aug. 28 after it added safety and security requirements, while some smaller experimental runs remained paused.

OpenAI announced the GPT-6 Astra launch on Sept. 3 and classified the model at the Critical cybersecurity capability threshold under its Preparedness Framework. The company said GPT-6 Astra was broadly deployed, although access to its most advanced cybersecurity capabilities would initially remain limited to a small group of testers.