OpenAI has reportedly walked away from a model because it would not reliably do what it was told.
A senior OpenAI executive told the Wall Street Journal that the company scrapped the model after it showed a poor aptitude for following instructions. The executive did not name the model or say when it was in development. No further detail on the decision process, testing, or scope of the problem has been disclosed. OpenAI has not issued its own public statement on the matter.
This is a different kind of failure than the usual "the model hallucinated" story. An assistant that will not follow orders is not just underperforming, it is unpredictable, and unpredictability is the exact thing safety teams are paid to catch before release. That an executive framed this as a safety call, not a performance one, suggests OpenAI is drawing a harder line between models that are merely weak and models that are simply unmanageable.
We have no numbers, no model name, and no timeline here, just an executive's characterization to a reporter. Take the safety framing as OpenAI's chosen story, not an independently verified account of what actually went wrong.