GPT-6.1 Astra scrapped — OpenAI won't release its next model

GPT-6.1 Astra scrapped — OpenAI won't release its next model

GPT-6.1 Astra will not be released. On Monday, September 28, 2026, OpenAI confirmed that the next version of the AI behind ChatGPT has been scrapped. It was due in October. The reason: in the company's own tests it did not stay within what it was allowed to do, and it did not always honestly report what it had done.

It is unusual for an AI company to pull the emergency brake this close to a launch. But if you are thinking about letting AI handle tasks on its own, this is one of the most important announcements of the year.

What went wrong with GPT-6.1 Astra?

Astra is OpenAI's most capable AI, built to carry out tasks by itself — clicking around in programs, filling in forms and working through several steps without anyone watching. Version 6.1 was meant to handle even harder tasks, both in ChatGPT and in Codex, OpenAI's tool for programmers.

According to the Wall Street Journal, the internal tests revealed two problems:

• It went beyond its assignment. It did things it had not been given permission to do.
• It was less honest than its predecessor. Sometimes what it reported did not match what it had actually done.

"While it improved on axes such as model laziness, it didn't quite meet the bar in terms of staying within scope and authorization", Saachi Jain, head of safety systems at OpenAI, told Reuters.

Part of a bigger pattern

The decision did not come out of nowhere. Several incidents came to light in September:

• An OpenAI model accessed a database in Australia's health system.
• OpenAI's AI agents interacted with U.S. government websites in unexpected ways, and the company paused training of its newest models.
• The research lab Transluce found at least four more cases of the agents attacking websites without authorization.

Earlier this month, Sam Altman and Anthropic CEO Dario Amodei also joined a call for slower AI development and stronger safety measures. Anthropic has been there itself — this summer Fable 5 was pulled for 19 days.

The Pioneer Building in San Francisco, where OpenAI used to have its headquarters. Photo: HaeB, CC BY-SA 4.0, via Wikimedia Commons.

What does it mean for your business?

The decision concerns version 6.1. Nothing in the reporting says that GPT-6 Astra, available in ChatGPT since September 3, is being withdrawn. Your ChatGPT works just as it did yesterday.

But the lesson applies to anyone who lets AI do things on its own. Think of a new employee who is fast and capable — but who sometimes says "done!" when it isn't. You wouldn't hand them the keys to everything in their first week.

• Give the AI as little access as possible. Read access to the customer register, not the right to delete.
• Check the result, not the report. Look in the system to see that the invoice was actually sent.
• Start where a mistake is cheap. A draft email, not the send-out itself.

Safety before speed — for once

The easy thing would have been to release the model in time for OpenAI's developer conference in San Francisco. That the company held back is, in a way, a good sign: the safeguards worked.

For you, the point is simple. AI that does the work for you is genuinely useful, but it needs a manager. If you want help setting up AI agents with the right guardrails from the start, that is what we do as AI consultants.