OpenAI’s new GPT-Live voice models change more than the sound of ChatGPT. They change the conversation loop itself: the system can keep listening while it speaks, decide whether a pause means “wait” or “answer,” and hand a difficult request to a separate reasoning model without ending the spoken exchange.
That architecture makes GPT-Live more interesting than a routine Voice update. It also makes the limits easy to misunderstand. GPT-Live is available in ChatGPT now, but not yet in the API; it does not launch with video or screen sharing; and its audio retention differs from the older Standard mode. Here is the practical map, checked against OpenAI’s official materials on 2026-07-20.
The important change is full-duplex conversation
Older voice assistants usually behave like walkie-talkies. You speak, the system decides your turn has ended, it processes the request, and then it talks. Even an audio-native model can feel rigid if silence is the main signal for turn-taking: a thinking pause can be mistaken for completion, while background noise can trigger an interruption.
GPT-Live is full-duplex, meaning it processes incoming audio while generating outgoing audio. OpenAI says it makes interaction decisions many times per second: speak, listen, pause, interrupt, or invoke a tool. In practice, that should make three ordinary behaviors less awkward: interrupting an answer, thinking aloud with pauses, and giving short acknowledgments without surrendering your turn.
The distinction is architectural, not a promise of perfect turn-taking. OpenAI’s ChatGPT Voice help page warns that overlapping speech, background noise, network conditions, and microphone settings can still affect what the model hears. It is built primarily for one-on-one conversation, not a room full of speakers.
One model talks while another model works
GPT-Live also separates the conversational layer from deeper reasoning. The voice model handles timing and spoken interaction, while a frontier model can perform search or reasoning in the background. At launch, Instant conversations and GPT-Live-1 mini use GPT-5.5 Instant behind the scenes; Medium and High use GPT-5.5 Thinking at the corresponding reasoning effort, according to OpenAI’s launch announcement.
This split matters because latency and reasoning depth pull in opposite directions. A model that pauses for every hard question feels unresponsive; a model optimized only for immediate speech gives shallow answers. Delegation lets the interface remain conversational while harder work proceeds elsewhere.
Do not read that as background autonomy without boundaries. The launch is a ChatGPT Voice feature, not a general-purpose voice-agent API. OpenAI says API access is planned but is not available at launch. GPT-Live also does not initially work with custom GPTs, Work, Codex, connected apps, or plugins.
Live, Advanced, or Standard: which mode fits?
ChatGPT now has three Voice paths, and the newest one is not automatically the right one for every task.
- Use Live for conversation. It is the best fit when interruptions, pacing, language practice, brainstorming, or hands-free back-and-forth matter. It can use web search and memory and can respond to text and images added in the same chat.
- Use Advanced when you need a camera or screen sharing. Live does not support either at launch. Eligible mobile subscribers can keep using those features through Advanced Voice Mode.
- Use Standard when you want a turn-by-turn exchange or different audio handling. Standard transcribes a completed turn before generating a response. OpenAI says Standard audio is deleted after transcription unless the user opted to share audio for model improvement.
There is another reason to keep the modes straight: a Voice transcript is not a verbatim recording. Fast speech, overlap, or background noise can produce differences between what was said and what appears in chat. If the output must become an exact meeting record, dictation or a dedicated transcription workflow remains the safer choice.
Plan limits are measured by time
Live usage uses a rolling 24-hour window, and OpenAI says the limits may change. At the time of checking, the official help page lists:
- ChatGPT Pro at $200 per month: unlimited GPT-Live-1 access.
- ChatGPT Pro at $100 per month: up to 12 hours of Instant, 12 hours of Medium or High, and 24 hours of GPT-Live-1 mini.
- ChatGPT Go and Plus: up to one hour of Instant, one hour of Medium or High, and two hours of GPT-Live-1 mini.
- ChatGPT Free: limited GPT-Live-1 mini access in each rolling 24-hour period.
A single Live conversation can last up to two hours. Availability can also depend on plan, region, workspace, and app version, so these figures should be treated as a dated snapshot rather than a permanent entitlement.
The privacy detail worth checking before you talk
Live and Advanced audio clips are stored with the chat transcript and retained for 30 days, according to OpenAI’s help documentation. Deleting the chat starts deletion of the associated clips within 30 days, subject to stated security, safety, and legal exceptions. Archiving a chat does not delete the clips.
OpenAI says it does not use audio or video clips for model training unless the user chooses to share them. Transcripts and other files may be used depending on plan and data-control settings when “Improve the model for everyone” is enabled. Anyone discussing sensitive personal or company information should review those controls before treating Voice like an informal private room.
The GPT-Live system card describes another voice-specific layer: safety checks run as the conversation unfolds and can steer, interrupt, or end a response. It also reports slight regressions in two adversarial evaluation categories, emotional reliance for GPT-Live-1 and sexual content for GPT-Live-1 mini, though OpenAI says neither difference was statistically significant. That is a useful reminder that a more natural voice can feel more trustworthy without becoming error-free.
What GPT-Live changes — and what it does not
GPT-Live’s real advance is coordination. Listening, speaking, deciding when to yield, and delegating difficult work now happen in parallel rather than as a visible chain of separate stages. That should make Voice more useful for thinking aloud, tutoring, translation, and hands-free research.
It does not remove the need to verify important answers, choose the right mode, or understand what is stored. It is also not yet the developer platform implied by the architecture; the API remains a future release. For now, the sensible test is small and personal: try Live on a conversation where pauses and interruptions matter, switch back to Advanced when the camera matters, and use text or transcription whenever exact wording matters more than conversational flow.