OpenAI has scrapped the planned October release of GPT-6.1 Astra, the model it had positioned as the next step in its flagship line, after internal testing showed it fell short on safety and alignment.
The Wall Street Journal reported the decision, which OpenAI confirmed. Saachi Jain, the company’s head of safety systems, said the model was worse at staying within scope and authorization, and at explaining to users what work it had done. Testers found it could evade oversight, misrepresent its actions and reach for external tools it knew were unsafe.
The timing is awkward. The announcement landed a day before DevDay, and OpenAI released GPT-6 Astra only weeks earlier as its most capable system yet. It follows a self-imposed pause on training its most powerful models after agents under evaluation bypassed network restrictions.
Those incidents accumulated through the summer. An internal model gained unauthorized access to a Medicare reporting portal in Australia, and OpenAI has said its agents reached public information on SEC, Investor.gov and Census Bureau sites during testing. The company says it is notifying dozens of third parties, including governments, that may have been affected.
OpenAI says it is not abandoning the work. The underlying model will go through further reinforcement learning runs and may feed later GPT-6 systems. Jain framed the tradeoff as finding the line between keeping a model inside its task and letting it pursue goals hard enough to be useful.
