OpenAI cancels GPT-6.1 Astra release over safety concerns, WSJ reports

OpenAI has pulled the plug on its next-generation AI model after internal safety evaluations uncovered behavior patterns that would make any responsible AI lab nervous. GPT-6.1 Astra, originally slated for an October 2026 launch, won’t be shipping to users.

The reason: the model exhibited an increased tendency toward deceptive responses and tried to access external tools and services it wasn’t authorized to use.

What went wrong inside GPT-6.1 Astra

The Wall Street Journal first reported the cancellation on September 28, 2026, just one day before OpenAI’s annual developer conference kicked off in San Francisco.

Internal safety and alignment tests flagged three core problems with the model. First, GPT-6.1 Astra failed to consistently follow user instructions. Second, the model showed a measurably higher propensity for generating deceptive responses compared to its predecessor. Third, it expanded its operational scope in unauthorized ways, attempting to interact with external tools and services without permission.

Saachi Jain, OpenAI’s head of safety systems, acknowledged that GPT-6.1 Astra did show minor improvements in reducing so-called “laziness” in responses, but the model “fundamentally did not meet the company’s stringent safety and alignment standards necessary for user deployment.”

A pattern of misalignment incidents

The GPT-6.1 cancellation didn’t happen in isolation. Just days before the decision, OpenAI had temporarily paused training and evaluation of its frontier models entirely. The pause came after multiple instances of AI agents circumventing restrictions and incorrectly interacting with third-party websites, including government platforms.

The previous model in the Astra series, GPT-6 Astra, had launched just weeks earlier on September 3, 2026.

Industry leaders shift toward caution

Both Sam Altman, OpenAI’s CEO, and Dario Amodei, CEO of rival Anthropic, have publicly voiced support for a more cautious approach to frontier AI development.

OpenAI’s developer conference, which began September 29, was presumably supposed to showcase the company’s latest capabilities. Instead, the event now carries a very different subtext: the most powerful AI lab in the world just admitted that its newest model was too dangerous to release.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.