OpenAI cancels new AI model launch last-minute over safety concerns: All details

OpenAI puts a last-minute pause on the rollout of its new frontier AI model, GPT-6.1 Astra, due to safety fears. The release was reportedly planned for October 2026, but it has now been cancelled after internal testing found that the model did not meet the company's safety and alignment standards, OpenAI confirmed on September 28. This is a rare instance in which an AI company has halted the release due to risks.

According to The Wall Street Journal report, the GPT-6.1 Astra model was designed to be integrated into ChatGPT and Codex to manage complex tasks with less human interaction. However, the model was not always transparent about its own actions, and this happened more often than with its predecessor. As a result, OpenAI decided to pause the release for October.

Must read: OpenAI agents reportedly breached US government websites during safety tests: Here's what happened

Saachi Jain, head of safety systems at OpenAI, said, “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.”

“We want to make sure our model development is safe no matter whether that's in the company, or when we ship it to users. But when we ship it to users, we have an extremely high bar in terms of safety and alignment,” Jain added. The decision comes days before the DevDay developer conference in San Francisco, where it is said to make more than 20 announcements, including new models, features, and updates during the event.

Must read: Meta Muse AI allegedly shared user’s address, finalised a sale deal without approval: Buyer showed up at 9:15 PM

The company is facing backlash over its safety guardrails after high-profile incidents in which its AI breached government websites. Last week, an OpenAI agent reportedly breached several government websites and systems, including Services Australia, NSW Bureau of Crime Statistics and Research, Victorian Department of Health, and Australian Institute of Health and Welfare.

Later, OpenAI apologised and admitted that it should have handled its response better. The company also assured that it will develop a practical approach for developers and governments to identify and disclose such AI incidents.

Previously, OpenAI also revealed that its AI system hacked the Hugging Face system and accessed credentials during internal testing. This has led the company to carefully test AI models before putting them in the hands of users.