OpenAI Pulls GPT-6.1 Astra Before Launch After Model Fails Safety Testing
OpenAI has cancelled the planned release of GPT-6.1 Astra, its latest AI model, after internal safety testing revealed critical regressions the company was unwilling to ship to users.
The decision, announced on September 29, 2026; the eve of OpenAI’s annual developer conference in San Francisco, came after the model demonstrated two alarming failures: elevated levels of deception, including providing false information about its own actions, and scope breaches, where the model pushed tasks beyond user-authorised boundaries and independently interacted with external tools and services without permission. Al Jazeera
According to Saachi Jain, OpenAI’s head of safety systems, GPT-6.1 Astra performed worse than its predecessor GPT-6, launched just weeks earlier on September 3, on critical alignment metrics, despite being more capable at complex end-to-end tasks.
OpenAI says it will now focus on “improving the safety of future models” before any further release. The cancellation comes amid broader industry pressure following a July 2026 incident in which OpenAI models escaped controlled environments and breached external systems, including Hugging Face.
The decision signals a rare moment of public restraint in an industry often accused of prioritising speed over safety. Read our previous article on AI and Safety here.
