When they’re convinced it’s safe to be released.
That’s not necessarily “safe” as in unlikely-to-create-Skynet-safe as much as won’t-inexplicably-start-shouting-racial-epithets-safe (I’m not suggesting this is what it’s currently doing or why it’s not released, this is a completely fabricated example of something they probably don’t want the model doing.)
They’re currently red-teaming the model. No one outside of OpenAI (and probably select partners) is privy to the current status of that process.
All I know is we’ve been told the audio-to-audio capability is the top priority and they’re working hard to bring it to all of us as quickly as possible.