# GPT-Live

> Source: https://aiwiki.ai/wiki/gpt_live
> Updated: 2026-07-24
> Fact-checked: 2026-08-05
> Categories: AI Models, OpenAI, Speech & Audio AI, Voice AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "GPT-Live." aiwiki.ai, 24 Jul 2026. https://aiwiki.ai/wiki/gpt_live
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

GPT-Live is a family of full-duplex [voice models](https://aiwiki.ai/wiki/voice_ai) developed by [OpenAI](https://aiwiki.ai/wiki/openai) for continuous spoken interaction. OpenAI released GPT-Live-1 and GPT-Live-1 mini on July 8, 2026, as the models behind a new Live option in [ChatGPT](https://aiwiki.ai/wiki/chatgpt) Voice. GPT-Live-1 is intended for paid users, while GPT-Live-1 mini is the lower-cost version used for Free accounts. Both can process incoming speech while producing speech, instead of waiting for a clearly bounded user turn to end.[1][2]

GPT-Live is both a model family and part of an orchestrated product. The voice model manages listening, speaking, interruptions, pauses, and short conversational responses. For questions that require web search or more computation, it can delegate work to a separate frontier model and continue the conversation while that work runs. At launch, GPT-Live-1 in Instant mode and GPT-Live-1 mini delegated to [GPT-5.5](https://aiwiki.ai/wiki/gpt-5.5) Instant. The Medium and High settings used GPT-5.5 Thinking with medium or high reasoning effort.[1]

## Release and product role

OpenAI announced and began the global consumer rollout on July 8, 2026, across ChatGPT.com and the iOS and Android applications. The launch made GPT-Live-1 the default Live model for Go, Plus, and Pro users and GPT-Live-1 mini the default for Free users. The older [Advanced Voice Mode](https://aiwiki.ai/wiki/advanced_voice_mode) and Standard Voice remained available as separate options. Live did not support Business, Enterprise, or Edu workspaces on launch day, but OpenAI's help documentation later added early access for eligible Business, Enterprise, Edu, and Healthcare workspaces, subject to administrator settings.[3][4]

The three ChatGPT Voice options use different interaction designs:

| Voice option | Basic design | Role after the GPT-Live launch |
|---|---|---|
| Standard | A cascade of [speech recognition](https://aiwiki.ai/wiki/speech_recognition), a text model, and [text-to-speech](https://aiwiki.ai/wiki/text_to_speech_ai) | Turn-by-turn fallback and legacy option |
| Advanced | Native audio input and output in one model, but organized as discrete turns | Previous real-time experience; retained for features such as video and screen sharing |
| Live | Continuous, full-duplex audio with background delegation | New default voice experience, using GPT-Live-1 or GPT-Live-1 mini |

GPT-Live is not the same product as the developer-facing [OpenAI Realtime API](https://aiwiki.ai/wiki/gpt_realtime). OpenAI said at launch that it planned to add GPT-Live-1 to the API and opened a notification form, but had not released API access by July 24, 2026.[1][5] An August 3, 2026 engineering post continued to describe the GPT-Live API as upcoming.[8] The API model catalog continued to list the separate GPT-Realtime family for developers. Consequently, GPT-Live had no public API model identifier, token price, context-window specification, or rate-limit table at that date.

## Continuous interaction and delegation

Earlier voice systems treated conversation as a sequence of requests. Standard Voice waited for speech to end, transcribed it, sent the transcript to a language model, and synthesized the written response. Advanced Voice removed the transcription and synthesis handoffs, but still depended on detecting when one turn had finished before beginning the next. A pause or background sound could therefore be mistaken for the end of a turn.[1]

GPT-Live continuously processes incoming audio while it generates output. OpenAI says the model makes interaction decisions many times per second, including whether to speak, keep listening, pause, stop, or invoke a tool. This permits conversational backchannels such as short acknowledgments, interruption while the model is talking, waiting through a user's pause, and live translation. It also changes turn detection from a single silence threshold into an ongoing model decision.[1]

Delegation separates conversational timing from deeper work. GPT-Live can acknowledge a request and keep interacting while a second model searches or reasons in the background, then introduce the result when it is ready. This arrangement can reduce dead air, but it also means that the perceived intelligence of the ChatGPT experience is not a measurement of GPT-Live alone. OpenAI's system card states that its evaluations include delegation because that matches the deployed product.[1][2]

OpenAI did not publish a single end-to-end latency number at launch. An independent test by Agora Media Lab measured the consumer ChatGPT application rather than direct model inference. On one iPhone 13, in English, with 30 trials per condition, it found about 1.1 seconds from the end of user speech to the first audible response. GPT-Live took about 1.4 seconds to stop after a deliberate interruption, roughly 0.5 seconds longer than Advanced Voice, but it was much less likely to stop for background speech. Under simulated 10% uplink packet loss, GPT-Live's median response delay increased by 314 milliseconds, compared with 2,448 milliseconds for Advanced Voice. The test was a launch-week snapshot from a communications vendor, not a broad independent benchmark.[7]

## System architecture

On August 3, 2026, OpenAI published an engineering post by Justin Uberti and Zahan Malkani describing the serving system built for GPT-Live, which it called its third-generation voice system. In an accompanying post on X, OpenAI said it "rebuilt the voice stack from client to model" so that continuous audio flow, deeper reasoning, and tool use do not interrupt the conversation. According to the post, earlier voice systems relied on a small turn-detector model to guess when a user had finished speaking before the larger model could begin inference; the full-duplex voice model removes the turn detector from the audio path and makes that decision itself. OpenAI said it reworked model inference, context management, and media transport over roughly six months to keep speech flowing continuously.[8][9]

The architecture separates the real-time media path from application logic. Audio streams between the client and the voice model on a dedicated fast path over WebRTC, while delegation, tool use, and other application work sit behind an asynchronous remote procedure call boundary, so a slow tool call can delay its own result but cannot stall the audio. OpenAI rewrote the media frontend and inference logic in Go, replacing a Python asyncio implementation, and reported that the new system's 95th-percentile frame-delivery smoothness matched the previous system's median. An architecture diagram published with the announcement shows the delegation path running from a media frontend through an application server to a GPT-5.5 model instance with search, code, and retrieval tools.[8][9]

Because a voice session is stateful and can remain active for a long time, the system uses a handoff mechanism: it warms a replacement model instance, prefills it with the current session context, runs both instances in parallel, and cuts over when the replacement is ready. The same mechanism performs context compaction off the live path when a long conversation approaches the model's context limit, avoiding the delay of rebuilding the [key-value cache](https://aiwiki.ai/wiki/kv_cache) mid-conversation. To reduce delegation latency, the application server creates and prefills a frontier-model inference session when the voice session starts and keeps it available with session affinity and prompt caching. The application server also converts the continuous, overlapping audio into discrete user and assistant messages for ChatGPT's interface, analytics, and safety systems, maintaining a provisional view for display and a finalized record for logging.[8]

For session startup, OpenAI developed the WebRTC Abridged Roundtrip Protocol (WARP), a set of backward-compatible protocol changes that it says reduce media and data-channel startup from six network round trips to one, including piggybacking the DTLS handshake over ICE, using DTLS 1.3, and pre-negotiating the SCTP handshake and data channels. OpenAI is advancing WARP as open specifications through the IETF TSVWG working group, and support has been added to the libwebrtc and Pion implementations. A separate mechanism OpenAI calls Instant Connect negotiates WebRTC signaling parameters ahead of time; together, OpenAI says, Instant Connect and WARP allow a client to start a session with a single UDP packet. Before launch, OpenAI ran a silent test that routed a gradually increasing share of production ChatGPT Voice sessions to the new system in read-only mode while Advanced Voice Mode continued serving users. The company said the architecture also powers voice-driven agent coordination on the ChatGPT desktop application and will underpin the planned GPT-Live API.[8]

## Modalities and ChatGPT features

The core GPT-Live interaction is audio input and audio output, with simultaneous listening and speaking. In ChatGPT, spoken responses also appear as streamed text. Users can type messages or attach images in the same chat, and Live can return supported visual cards for information such as weather, markets, sports, or maps. These are capabilities of the deployed ChatGPT experience, and OpenAI has not disclosed which parts run directly in GPT-Live rather than through product-level routing or delegation.[1][3]

The launch version did not accept live video or screen sharing. Those inputs remained available in Advanced Voice for eligible mobile subscribers. Live also initially excluded connected applications and plugins, was unavailable inside custom GPTs, and could not independently perform image generation, data analysis, or custom actions in a custom-GPT voice conversation. OpenAI's help page later documented a separate Voice in Work and Codex experience on the ChatGPT desktop application, where the voice interface can coordinate agents using the tools and permissions supplied by those products.[3]

OpenAI provides nine preset ChatGPT voices for Live. The system is designed for conversation rather than voice cloning, and its safeguards are intended to prevent imitation of a real person's voice. OpenAI has not released the model weights, parameter counts, context limits, knowledge cutoff, model-level architectural details beyond the full-duplex and delegation design, or a list of native input and output formats outside the ChatGPT product, though an August 3, 2026 engineering post later described the serving system built around the model.[1][2][8]

## Access and usage limits

GPT-Live did not have standalone API pricing at launch because it was not available through the API. Consumer access was bundled into ChatGPT subscriptions and governed by rolling usage limits rather than per-token billing. OpenAI's July 24 help documentation listed the following allowances, while warning that limits can change:[3]

| Plan | Documented Live allowance |
|---|---|
| Free | Limited GPT-Live-1 mini use in each rolling 24-hour period |
| Go and Plus | Up to 1 hour of GPT-Live-1 Instant, 1 hour of Medium or High, and 2 hours of mini |
| Pro at the help page's $100 tier | Up to 12 hours each of GPT-Live-1 Instant and Medium or High, plus 24 hours of mini |
| Pro at the help page's $200 tier | Unlimited GPT-Live-1 access |
| Business | 1 hour each of Instant and Medium or High; further use at 5 credits per minute |
| Enterprise, Edu, and Healthcare on flexible pricing | 5 credits per minute |

A single Live conversation can last no more than two hours. OpenAI separately listed Voice in Work and Codex at approximately 6 credits per minute for Business and Enterprise workspaces on flexible pricing, with delegated tasks charged against the relevant product's normal usage pool.[3] These plan entitlements are product pricing rather than a price for running GPT-Live as a standalone model.

## Evaluation

OpenAI created head-to-head human evaluations using matched conversations lasting 5-10 minutes. It reported that both GPT-Live models were strongly preferred to their Advanced Voice predecessors for overall preference, turn-taking, interruptions, conversational flow, and perceived naturalness. The company did not publish the sample size, preference percentages, or full evaluation protocol on the launch page.[1]

OpenAI also reported that GPT-Live-1 outperformed Advanced Voice on [GPQA](https://aiwiki.ai/wiki/gpqa), [BrowseComp](https://aiwiki.ai/wiki/browsecomp), and an internal variant of the tau3-Voice Telecom task. These results included the deployed delegation system: Instant and mini used GPT-5.5 Instant, while Medium and High used GPT-5.5 Thinking. They therefore test the combined voice-and-reasoning product rather than isolating the voice model's own scientific knowledge or search ability.[1][2]

The system card reports separate voice-native safety evaluations. On difficult production prompts selected from cases where Advanced Voice had not responded ideally, GPT-Live-1 was equal to or better than its predecessor in most measured categories. Its emotional-reliance score fell from 0.88 to 0.82, and the mini model's sexual-content score fell from 0.97 to 0.95; OpenAI said neither difference was statistically significant. On a synthetic adversarial set covering sexual content, illicit behavior, mental health, personal data, emotional reliance, self-harm, hate, and gore, both GPT-Live models matched or exceeded their predecessors in every reported category. OpenAI cautioned that neither set was prevalence-weighted and that the scores should not be treated as estimates of safety in ordinary use.[2]

## Safety work

OpenAI trained GPT-Live with data from publicly available internet sources, licensed or partnered sources, and material provided or generated by users, human trainers, and researchers. The system card describes filtering intended to reduce personal information and harmful or sensitive training material, but does not give dataset sizes or a detailed data inventory.[2]

The safety evaluation program included audio prompts from users who had opted to share voice interactions, with permission and deletion checks, eligibility filtering, de-identification, and personal-information scrubbing. OpenAI also generated synthetic spoken prompts for rare or adversarial cases. Internal and external red teamers tested the models across languages for self-harm, emotional reliance, manipulation, child-coded voices, impersonation, speaker identification, and audio-specific perturbations.[2]

Runtime safeguards examine inputs and outputs while a conversation unfolds. They can steer or interrupt a response, play a spoken safety message, display support resources, or end a higher-risk conversation. OpenAI also adapted its self-harm support flows for voice, added age-appropriate behavior for teen accounts, and began post-launch monitoring focused on emotional reliance.[1][2]

Under OpenAI's [Preparedness Framework](https://aiwiki.ai/wiki/preparedness_framework), its Safety Advisory Group concluded that GPT-Live-1 and GPT-Live-1 mini, without delegation, could not plausibly reach the High threshold for biological and chemical risk, cybersecurity, or AI self-improvement. Delegated work inherits the safeguards of the model that performs it. The standalone voice models did not have code execution or broad independent tool access at launch, and OpenAI did not run AI self-improvement evaluations because the company found them less capable than GPT-5.5 Thinking on several intelligence tests.[2]

## Limitations

GPT-Live remains sensitive to conversational and acoustic conditions. OpenAI says it is designed mainly for one-on-one use and is not optimized for several people speaking together. Overlapping speech, background noise, network quality, microphone settings, and long pauses can still cause interruptions or missed speech. Transcripts may differ from what either party actually said. Agora's test also observed four unsolicited responses to background speakers in 30 noisy-room trials, a speaker-attribution failure not present in the two older modes in that test.[3][7]

Language quality varies. OpenAI says it optimized the models for some of ChatGPT's most popular languages but did not publish the list. It warns that other languages may have non-native accents or gaps in fluency. At OpenAI's launch demonstration, TechCrunch described the Hindi translation as having a strong American accent and wording that sounded unnatural and overly formal.[1][6]

The lack of API access limits independent reproduction. Researchers cannot directly control model settings, isolate GPT-Live from ChatGPT's delegation and interface, or compare usage cost under a fixed API workload. OpenAI's capability evaluations are developer-reported, and several measure the whole deployed system. Model size, context limits, training-compute details, and complete architectural specifications remain undisclosed.[1][2][5]

## References

1. OpenAI. "Introducing GPT-Live." July 8, 2026. https://openai.com/index/introducing-gpt-live/
2. OpenAI. "GPT-Live System Card." July 8, 2026. https://deploymentsafety.openai.com/gpt-live
3. OpenAI Help Center. "ChatGPT Voice." Updated July 24, 2026. https://help.openai.com/en/articles/20001274-gpt-live-faq
4. OpenAI Help Center. "ChatGPT Release Notes." July 8, 2026 entry. https://help.openai.com/en/articles/6825453-chatgpt-release-notes
5. OpenAI. "GPT-Live-1 in the API." Accessed July 24, 2026. https://openai.com/form/gpt-live-1-in-the-api/
6. Ivan Mehta. "OpenAI releases new voice models for more natural live conversations." TechCrunch, July 8, 2026. https://techcrunch.com/2026/07/08/openai-releases-new-voice-models-for-more-natural-live-conversations/
7. Hermes Frangoudis. "OpenAI Didn't Publish GPT-Live's Latency. So We Measured It." Agora Media Lab, July 10, 2026. https://prod.agora.io/es/blog/openai-didnt-publish-gpt-lives-latency-so-we-measured-it
8. Justin Uberti and Zahan Malkani. "How we built a realtime system for responsive voice AI in six months." OpenAI, August 3, 2026. https://openai.com/index/continuous-voice-interaction-with-gpt-live/
9. OpenAI (@OpenAI). "GPT-Live can listen while it speaks." X post, August 3, 2026. https://x.com/OpenAI/status/2084378415818579975