AI News

GPT-Live Now: ChatGPT Voice Modes, Limits, and Retention

Choose ChatGPT Voice Live, Advanced, Standard, or Dictation using current limits, feature support, retention, and fallback paths.

  • #gpt-live
  • #chatgpt
  • #voice-ai
  • #openai
  • #multimodal

Choose Live for natural interruption and back-and-forth conversation, Advanced when you need mobile camera or screen sharing, Dictation when you want to edit words before sending them, and Standard for a turn-by-turn voice exchange. Do not use a Voice transcript as an exact meeting record. Check retention settings and whether Live is available on your account before moving a sensitive workflow into it.

This is a mode-selection problem, not a contest in which the newest option always wins. Each path has a different interaction model, feature boundary, and recovery route.

The current Voice choices

GPT-Live uses a full-duplex architecture: it can listen and speak at the same time. Full-duplex is closer to a telephone conversation than a walkie-talkie, where one side must finish before the other starts. That design supports interruptions and short acknowledgments, but it does not eliminate errors caused by overlapping speech, background noise, long pauses, network conditions, or microphone settings.

ItemVerified valuePrimary sourceChecked
LiveGPT-Live-1 on paid plans and GPT-Live-1 mini on Free; supports web search, Memory, supported widgets, text, and images in the same chatChatGPT Voice FAQ2026-08-24
AdvancedPrevious real-time Voice experience; supports video and screen sharing for eligible iOS and Android subscribersChatGPT Voice FAQ2026-08-24
StandardTurn-by-turn Voice that transcribes speech before generating a responseChatGPT Voice FAQ2026-08-24
Not initially supported in LiveVideo, screen sharing, connected apps, and plugins; Live is also unavailable with custom GPTsChatGPT Voice FAQ2026-08-24
Maximum Live conversationTwo hoursChatGPT Voice FAQ2026-08-24
Simultaneous Voice conversationsOneChatGPT Voice FAQ2026-08-24
Supported surfacesEligible web, iOS, Android, and Desktop experiences; availability varies by plan, region, workspace, and app versionChatGPT Voice FAQ2026-08-24
GPT-Live developer APIThe launch page says API access is planned; no generally available generation model ID or price was foundGPT-Live announcement2026-08-24

The FAQ also describes Voice in Work and Codex on ChatGPT Desktop. It can start, prioritize, interrupt, or redirect tasks and coordinate agents on supported Windows and macOS desktop experiences. Standalone Work or Codex Voice is not available on web or mobile, although paired iOS remote access may be supported. Its allowance and pricing are separate from ordinary Voice in Chat.

Decision table: choose the interaction, then keep a fallback

What you needChooseWhyDo not use it whenHow to returnBasis and checked date
Thinking aloud, language practice, or a conversation with interruptionsLiveSimultaneous listening and speaking better fits conversational turn-takingThe task requires several speakers or a verbatim recordSwitch to Standard under Settings > Voice and verify important details in textVoice FAQ, 2026-08-24
Mobile camera input or screen sharingAdvancedThose capabilities remain on eligible iOS and Android subscriptionsYou need a desktop-only workflow or the account is ineligibleStop sharing and attach a bounded image to a normal chat, or return to textVoice FAQ, 2026-08-24
Reviewing and correcting speech before it becomes a promptDictationIt turns one recording into editable text before submissionImmediate spoken back-and-forth is the primary needEdit or discard the draft, then start a Voice conversation if neededVoice FAQ, 2026-08-24
Turn-by-turn speech without retaining Live audio clipsStandardIt transcribes the completed turn before generating a response, then normally deletes the audioNatural interruption is requiredSwitch back to Live under Settings > VoiceVoice FAQ, 2026-08-24
An exact meeting, legal, or verbatim recordDo not use Voice as the system of recordVoice transcripts may differ from what was saidA dedicated recorder, transcription process, and human review are unavailableReturn to a workflow that preserves and checks the original recordingVoice FAQ, 2026-08-24
Embedding GPT-Live directly in an applicationDo not build on that assumption yetA generally available GPT-Live generation model ID and price were not foundThe plan treats a differently named Realtime API as the same product without checking requirementsKeep the test inside ChatGPT, or evaluate a published Realtime model as a separate productGPT-Live announcement, 2026-08-24
Discussing confidential informationPause until data controls and policy are confirmedLive and Advanced clips are stored with the transcript for 30 daysThe organization cannot confirm its allowed data path and OpenAI settingsEnd Voice and return to an approved text or internal channelVoice FAQ, 2026-08-24

Usage limits are rolling, not a midnight reset

Live usage is measured over a rolling 24-hour period. A rolling window follows usage over the preceding period rather than refilling one universal bucket at midnight. OpenAI says limits may change and ChatGPT will notify users when a limit is reached.

PlanPublished Live limitPrimary sourceChecked
ChatGPT Pro, listed at $200/monthUnlimited GPT-Live-1 accessChatGPT Voice FAQ2026-08-24
ChatGPT Pro, separately listed at $100/monthUp to 12 hours of GPT-Live-1 with Instant, 12 hours with Medium or High, and 24 hours of miniChatGPT Voice FAQ2026-08-24
ChatGPT Go and PlusUp to one hour with Instant, one hour with Medium or High, and two hours of miniChatGPT Voice FAQ2026-08-24
ChatGPT FreeLimited mini access in each rolling 24-hour period; no fixed duration foundChatGPT Voice FAQ2026-08-24
ChatGPT BusinessUp to one hour of Instant and one hour of Medium or High; additional use costs five credits per minuteChatGPT Voice FAQ2026-08-24
Enterprise, Edu, and Healthcare on flexible pricingFive credits per minuteChatGPT Voice FAQ2026-08-24
Legacy Enterprise and EduUp to one hour of Live and two hours of miniChatGPT Voice FAQ2026-08-24

The same official FAQ lists “ChatGPT Pro” at both $200 and $100 without explaining the formal relationship between those entries in the text reviewed for this article. This article does not invent replacement plan names. Before changing a subscription for Voice alone, check the price and limit shown in Settings > Subscription on the account that will use it.

A single Live conversation can last no more than two hours even when the daily allowance is larger or described as unlimited. A session may also end at a usage limit, maximum session length, or context limit. The fallback is text or a new Voice conversation, not an assumption that the previous audio session will continue indefinitely.

Retention and deletion should be part of mode selection

Audio clips from Live and Advanced, and video clips from Advanced, are stored with the transcript in chat history and retained for 30 days. Deleting the chat starts deletion of associated clips within 30 days, subject to OpenAI’s stated security, safety, and legal exceptions. Archiving is not deletion; it removes the chat from the sidebar without deleting the chat or its clips.

Standard handles audio differently. It transcribes the audio before generating a response, then deletes the audio after transcription, including when transcription fails, unless the user has chosen to share audio for model improvement.

OpenAI says audio and video clips are not used for training unless the user chooses to share them or enables the relevant recording-sharing settings. Transcripts and other files may be used depending on the plan and Data Controls. A personal account, a Business workspace, and an Enterprise workspace should not be assumed to have identical settings.

Review Settings > Data Controls before speaking customer names, health information, credentials, private company information, or unpublished material. Natural conversation can feel more private than a typed form, but the product’s storage rules, not the interface’s tone, define the data boundary.

A reproducible account check

Use a non-sensitive test to verify what the current account can do.

  1. Update ChatGPT and open Settings > Voice.
  2. Record which of Live, Advanced, and Standard are actually available. The expected result is an account-specific mode list rather than an assumption based on a general article.
  3. Start a short Live conversation without personal data. Interrupt with a clear request and compare the displayed transcript with what was said.
  4. If screen sharing is needed, switch to Advanced on an eligible mobile device and share only a test screen. Hide notifications and private content first.
  5. For exact prompt wording, record with Dictation and confirm that the text can be edited before submission.
  6. Open Settings > Data Controls and identify the current model-improvement and audio or video sharing choices. The expected result is that the operator can explain what is enabled.
  7. If deleting the test chat, use Delete rather than Archive. Deletion cannot be undone, so review any needed transcript first.

This procedure tests availability, control placement, transcript behavior, and retention choices. It does not prove that Live will transcribe every accent, device, or noise condition accurately. Important names, dates, amounts, and commitments still need text confirmation.

Do not assume a GPT-Live generation API is available

The July 8 launch announcement says GPT-Live-1 and mini were rolling out in ChatGPT and that API access was planned. A July 31 update discusses SynthID on supported audio generated through ChatGPT Voice and the OpenAI API, plus verification access. The reviewed launch page and model catalog did not provide a generally available GPT-Live generation model ID, price, or onboarding procedure.

OpenAI separately publishes GPT-Realtime models. A similar use case or audio interface does not make a differently named model identical to GPT-Live. Developers should evaluate the published Realtime model’s inputs, outputs, tools, latency, retention, and price against their own requirements.

Until OpenAI exposes a clear GPT-Live generation product surface, keep ChatGPT Voice evaluation separate from API architecture. That avoids building a dependency on a launch promise or confusing provenance verification with generation access.

The practical answer

Use Live for one-to-one conversation, Advanced for eligible mobile video or screen sharing, Standard for turn-by-turn voice, and Dictation for editable spoken input. Use a dedicated recording and transcription process when exact wording matters. Check the account’s current Voice options and Data Controls, then return to another mode or text when the chosen path does not fit.

The reviewed official material did not establish a general GPT-Live generation API release date, model ID, or price; the formal distinction between the FAQ’s $100 and $200 Pro entries; or a fixed duration for Free users.

Official primary sources

  • Introducing GPT-Live - full-duplex architecture, ChatGPT rollout, and planned API access. Checked 2026-08-24.
  • ChatGPT Voice - modes, plan limits, supported features, retention, deletion, and Work or Codex Voice. Checked 2026-08-24.
  • OpenAI API Models - published API catalog and separately named Realtime models. Checked 2026-08-24.