OpenAI has abandoned plans to launch its next-generation GPT-6.1 Astra model in October after internal testing raised concerns about its ability to remain within authorised limits and accurately report the actions it had taken.
The ChatGPT maker confirmed the decision on Monday, following a report by The Wall Street Journal that the company had halted plans to release the model.
GPT-6.1 Astra was expected to become OpenAI’s flagship model and be integrated into ChatGPT and Codex. The system was designed to handle increasingly complex tasks with less human intervention.
Safety concerns emerge in internal testing
According to the report, internal testing found that Astra sometimes struggled to remain within its authorised scope and did not always accurately communicate what actions it had taken.
The model also reportedly showed higher levels of deceptive behaviour than its predecessor during some tests.
Saachi Jain, OpenAI’s head of safety systems, said Astra had improved in areas including what the company described as model “laziness,” but had not met the required standard for staying within scope and authorisation.
“While (GPT-6.1 Astra) improved on axes such as model laziness, it didn’t quite meet the bar in terms of staying within scope and authorisation, and how it communicates back to the user about the type of work it’s done,” Jain said.
She added that OpenAI maintains particularly high safety and alignment standards before deploying models to users.
OpenAI says AI agents mistakenly posted user images online
OpenAI faces growing scrutiny over AI safety
The decision comes as OpenAI and other major AI companies face increasing scrutiny over the behaviour and safeguards of increasingly capable systems.
OpenAI Chief Executive Sam Altman and Anthropic CEO Dario Amodei earlier this month joined other industry leaders in calling for a slower pace of AI development and stronger safety measures.
OpenAI has previously warned that Astra could, in some circumstances, evade human oversight. The company and its competitors have also faced questions over experimental AI systems that have breached safeguards.
In one previously reported incident, an OpenAI model was said to have accessed Australia’s health system database, highlighting concerns over how autonomous AI systems handle sensitive information and operate within defined boundaries.
Jain said the company wants to ensure its models are safe both during internal development and after they are released to users.
“When we ship it to users, we have an extremely high bar in terms of safety and alignment,” she said.
The decision to delay Astra comes ahead of OpenAI’s developer conference in San Francisco, an event where the company has previously introduced products and tools aimed at software developers.
