GPT-Live

The old ChatGPT voice took turns. You spoke, it waited for silence, then it answered — a walkie-talkie with one button. GPT-Live, rolling out to ChatGPT users worldwide on July 8, drops that boundary: it processes your incoming audio while it is still talking, and makes a fresh decision — speak, pause, interrupt, hand off a hard query — many times a second. The everyday tell is that it can murmur 'mhmm' while you are mid-sentence, the way a person does.

The predecessor was already a single speech-to-speech model, so this is not about a smarter voice so much as the removal of the turn. When you ask something that needs real work, GPT-Live quietly delegates it to the heavier GPT-5.5 model in the background and keeps the conversation going instead of falling silent — the design point being conversations that run half an hour, not a single exchange. OpenAI reports its testers 'strongly preferred' the new model over the old voice mode on turn-taking and flow; those are the company's own human evaluations, not yet independently checked.

Worth keeping in view: continuous, interruptible voice is not new — Google's Gemini Live already does it, and there is even an academic benchmark for the capability. The news is OpenAI shipping it as the default for roughly a billion people, the free-tier 'mini' version replacing what most users currently hear. The seam still shows: reporters at the launch caught the model speaking Hindi in a heavy American accent with a stiff, bookish tone. The turn boundary is gone; the last mile of sounding human is not.