What is it about streaming specifically that necessitates this? Am I missing something obvious?
Excuse is probably that classifiers for streaming are less robust.
It's easier to get a partial response for something like a CBRN topic
I thought this was indeed the case at release, but then they changed it to also be for non-streaming (completions). So either they reverted it back or it was a temporary bug during the early days of the model.
When did you last check?
I ran a test just now, and gpt-5 without streaming works without the biometric check.
Interesting! Makes one wonder why they limit streaming only. I guess to induce just enough friction then.