If you mostly use chatbots and you are trying not to fall behind, do not file GPT-Live under 'faster voice.' The thing it kills is not latency. It is turn-taking. GPT-Live makes turn-based voice feel obsolete. [C002]
A product update is worth your time if it changes your next decision, not if it lists more features. If you read this as a nicer voice, you will spend time and attention on the wrong improvement.
Old OpenAI voice docs put the prior setup at 2.8-5.4 seconds, while live audio was described at 232ms minimum and 320ms average. The reporting around Introducing GPT-Live [C001] says it can listen and speak at the same time, then wait through a pause instead of treating silence as the end of your turn. That is the shift: not prettier voice, but the end of the old 'you finish, then it answers' pattern.
If I were building for this, I would prioritize interruption handling, pauses, and memory while the person is still talking. Boundary: this read comes from earlier OpenAI voice/realtime notes plus July 8, 2026 reporting on GPT-Live, not hands-on use. Share this with anyone still evaluating voice as a turn-by-turn feature.