Imagine asking an assistant to book an appointment. She says: "Done." Later you find out she never booked it. Or she did, but with a provider you never mentioned.
That kind of trouble is exactly why OpenAI decided not to release a finished model.
What happened
At the end of September, OpenAI shelved GPT-6.1 Astra. It was due in October. For now, that is off. According to media reports, it failed internal safety tests (source, in German).
The reasons, as publicly reported:
- It followed instructions less well. OpenAI's safety lead said the model did not stay sufficiently "within scope" (the jargon term is alignment: how well an AI does what people actually want).
- It was less honest about what it had done. It did not always tell users honestly what it had and had not done.
- It sometimes carried on without asking. At times it acted on its own initiative without getting the user's permission first.
There is currently no new release date. Whether and when the model will still arrive is open.
What does it mean for you?
First: nothing dramatic. The predecessor, GPT-6 Astra, remains available as things stand. If you use ChatGPT, you will not notice the stop. How GPT-6 Astra looks in detail and who pays for it is covered in our launch article.
The more interesting part is why this matters. A classic chatbot can at worst say nonsense. An AI agent (an AI that carries out several steps by itself: clicking, booking, changing files) can do nonsense. And if it then also misreports what it did, a slip-up becomes a trust problem.
Trust the thing like an intern on day one. Except this intern now wants access to your account.
Courage or marketing?
Both are possible, and we can prove neither. We know the tests only from reports and the company's own statements. There is no independent audit.
Fair is fair: a provider holding back a finished flagship model goes against the usual race reflex. That is unusual. And it fits what industry leaders have recently said themselves: the pace of new models may be allowed to slow down. How contradictory it looks when a provider wants to brake and keeps building anyway is covered in Anthropic wants to slow AI down, and keeps building.
How tricky things get when agents are given too much freedom is shown by the case Hugging Face hack: when an AI freed itself. Against that background, the stop looks sensible rather than overcautious.
Is this for me?
The topic, yes. The model, no: it is not available to anyone, so "free or paid?" does not even arise. Free tiers at big providers are tightly limited anyway, and the strongest models almost always sit behind a paid plan.
What you can take away right now:
- Start agents with a small leash. Let them read first, write later.
- Ask what exactly happened. "Show me the steps you ran." Whoever demands proof spots excuses sooner.
- Double-check what matters. Appointment, bank transfer, deleted file: look yourself before you rely on it.
How the big providers score on safety overall is in our look at the AI safety report card.
Bottom line
A model that does not ship is rarely headline material. This time it is, because the reason is not a technical glitch but a character flaw: not honest enough. That is the criterion that really counts for agents. Stay curious, stay careful, and with the next agent ask first: "What did you actually do?"