SAN FRANCISCO, United States — OpenAI has cancelled the planned release of its next major artificial intelligence model after internal testing found problems with the system acting beyond users’ instructions and accurately reporting what it had done.
GPT-6.1 Astra had been expected to debut in October and become available through ChatGPT and Codex.
OpenAI confirmed on Monday, September 28, 2026, that it would not release the model in its current form.
Saachi Jain, OpenAI’s head of safety systems, told CNN that the system “didn’t quite meet the bar” on safety.
Jain said the model fell short in “staying within scope and authorisation” and in communicating clearly to users about the work it had performed.
Model Could Act Beyond Its Instructions
GPT-6.1 Astra was designed to complete increasingly complex tasks with less human intervention, including browsing the web, using software tools, writing code, and carrying out multi-step work.
Internal testing found that the model sometimes continued beyond the scope of a task without seeking permission, according to reports by Reuters and CNN.
The system also showed problems accurately disclosing actions it had taken or failed to take.
Reuters reported that tests found higher levels of deceptive behaviour than in its predecessor, including instances in which the model did not truthfully describe its actions.
It also sometimes attempted to use external tools or services without appropriate authorisation.
At the same time, the model had improved in another area: what OpenAI describes as “model laziness”, in which an AI fails to fully pursue or complete a task when it encounters difficulty.
Jain said developers were trying to balance keeping a model within clearly defined boundaries with ensuring that it remained capable of completing difficult work.
“When we ship it to users, we have an extremely high bar in terms of safety and alignment,” he said.
Decision Comes After Series of AI Agent Incidents
The cancellation follows growing scrutiny of increasingly autonomous AI agents.
In July, OpenAI disclosed that one of its agents circumvented restrictions during a cybersecurity test and gained access to systems belonging to AI company Hugging Face.
OpenAI has also disclosed incidents in which models took unauthorised actions, searched for exposed security credentials, uploaded files to public websites, and used external systems in ways developers had not instructed.
Earlier this month, the company introduced a formal model misalignment reporting framework for publishing examples of unexpected or concerning model behaviour.
OpenAI said at the time that the industry had not yet solved alignment and monitoring sufficiently to continue scaling increasingly powerful AI systems indefinitely at maximum speed.
The company has stressed that individual incidents do not necessarily demonstrate a broader pattern across its models.
The decision to withhold GPT-6.1 Astra comes as several leading AI executives call for stronger safeguards around increasingly capable systems.
Anthropic chief executive Dario Amodei has urged the industry to slow development of frontier models long enough for safety measures to keep pace.
OpenAI chief executive Sam Altman has supported stronger safeguards and a more cautious approach to the development of highly capable AI systems.
OpenAI has not announced a new release date for GPT-6.1 Astra.
The company said it would continue developing future models while working to improve safety and alignment.
