OpenAI delays GPT-6.1 Astra over deception concerns

A pause, with an exception

OpenAI had previously said it would pause training its most powerful models after the recent incidents. But according to The Wall Street Journal, GPT-6.1 Astra was not included in that group.

That makes this delay different from a broad retreat: OpenAI is holding back one model while seeking to understand its behaviour and derive safer versions from it. The company has described the issue as deception, but the announcement as reported here does not say what that means in practice or how it was measured.

The unanswered question

The summer incidents sharpened two distinct concerns: the prospect of uncontrollable superintelligence capable of self-improvement, and the risks posed by existing systems that are already difficult to control. The case for slowing development does not depend on the first scenario; the second is already part of the debate.

I think the more consequential detail is not that OpenAI delayed a model, but that its earlier training pause apparently did not cover Astra. That leaves the boundary of the pause unclear: which models qualify as powerful enough to stop, and who decides when a safety problem is serious enough to hold one back?

It is also unclear whether other AI labs will slow their releases. There appears to be some industry agreement on slowing AI development, but a single delay does not show how far that agreement reaches.