- GPT-Live has powered ChatGPT Voice since July 8, 2026, with GPT-Live-1 for eligible paid users and GPT-Live-1 mini for free users across supported web and mobile apps. [1][4] - Unlike Standard and Advanced Voice, GPT-Live continuously processes incoming audio and can listen while speaking, without a separate turn detector controlling the live audio path. [1][2] - OpenAI has not published an end-to-end latency benchmark. One third-party test found only a 205-millisecond median improvement over Advanced Voice, but substantially less variation between responses. [3] - As of August 4, OpenAI still described a dedicated GPT-Live API as upcoming. Developers can use separate GPT-Realtime models through the Realtime API.
[2][7] OpenAI’s GPT-Live is a deployed consumer product, not simply a research demonstration. The company began rolling it out inside ChatGPT Voice on July 8, and an engineering account published August 3 describes the production architecture built to carry continuous voice conversations at ChatGPT scale.[1][2]
The distinction matters because OpenAI discusses GPT-Live as both a model advance and a broader realtime platform. ChatGPT users can access it now, subject to plan, region, workspace and app availability. Developers cannot yet build directly on the same GPT-Live models through a generally available API, although OpenAI says one is coming.[2][4]
What changed from earlier ChatGPT voice modes #
Standard Voice follows a three-stage pipeline: speech recognition converts the user’s words to text, a language model writes an answer, and a text-to-speech system reads it aloud. That structure makes text models accessible by voice, but each handoff adds delay and can discard information carried by tone, pace or emphasis.[1][2]
Advanced Voice improved on that design by using a model that directly processes and generates audio. It reduced latency and produced more expressive speech, but still organized conversations into discrete turns. A turn detector watched for silence and decided when the user had finished. Brief s or background sounds could therefore trigger an answer too early.[1][2]
GPT-Live removes that separate detector from the live audio path. Its full-duplex voice model continuously receives audio while generating speech, making decisions many times per second about whether to listen, speak, , interrupt or invoke a tool. OpenAI describes it as its third-generation voice system.[1][2]
That does not mean one model performs every task. GPT-Live handles the immediate conversation, while a frontier model can work asynchronously on searches, tool calls or difficult reasoning. OpenAI’s published examples use GPT-5.5 for delegated work, while the launch announcement says the background model can be updated as newer frontier systems become available.[1][2]
OpenAI’s disclosures describe the system architecture, not the underlying neural architecture in detail. The company has not published GPT-Live’s parameter count, training compute, model dimensions or a technical explanation of how its audio representation works. Claims that it is “full-duplex” or “two-model” should therefore be read as descriptions of deployed behavior and system design, not a complete model specification.
How OpenAI keeps the conversation moving #
The production system is built around a continuous media loop rather than a series of independent requests. Audio frames stream into a stateful model session and generated speech streams back. Conversation storage, context compaction and deeper reasoning run away from that critical path so they do not stop the audio flow.[2]
For delegated tasks, OpenAI creates and prefills a frontier-model session when the voice conversation begins. It keeps that session available, uses prompt caching and maintains affinity with the same serving infrastructure. The voice model can offer a brief acknowledgment while the second model searches or reasons, then incorporate the result when it arrives.[2] This arrangement can make the system feel responsive before it has produced the substantive answer. A quick acknowledgment still counts as audio, even if the search behind it takes several more seconds. Any latency comparison therefore needs to distinguish the first audible response from the start and completion of the useful answer.
What third-party latency testing found #
OpenAI has not released a numerical end-to-end latency benchmark for GPT-Live. Agora Media Lab, operated by realtime communications vendor Agora, tested GPT-Live-1, Advanced Voice and Standard Voice on an iPhone 13 shortly after launch. Its protocol used prerecorded English speech, an artificial mouth and separate waveform recordings, with 30 trials per condition.[3]
Agora found that GPT-Live’s median time from the end of the user’s speech to its first audible output was only 205 milliseconds better than Advanced Voice. The larger measured improvement was consistency: Advanced Voice had a 2,318-millisecond 90th-percentile response time, while GPT-Live’s 90th percentile was 104 milliseconds above its own median. The standard deviation in response timing fell from 489 milliseconds to 104 milliseconds.[3]
That result measures the first audible sound, which can be an acknowledgment rather than the answer itself. The test did not separately measure the beginning or completion of a substantive response, limiting what it establishes about latency on searches and reasoning tasks.[3]
Consistency also came with a trade-off. GPT-Live took 498 milliseconds longer than Advanced Voice to stop speaking after a deliberate interruption. Agora attributed the delay to waiting for more evidence that the sound represented a genuine attempt to take the floor. GPT-Live reacted incorrectly to one of 60 false-interruption probes, compared with 22 for Advanced Voice. In another test, however, it answered background conversations not directed at it in four of 30 ten-second windows.[3]
Under 10% uplink packet loss, GPT-Live added 314 milliseconds of median delay, while Advanced Voice added 2,448 milliseconds. Agora cautioned that its results came from one device, account and location during launch week. It also sells realtime communications infrastructure and disclosed its commercial interest in network and interruption performance. The results are useful as an early controlled test, but they are not an independent, multi-lab benchmark.[3]
Availability, privacy and safety boundaries #
GPT-Live-1 is the default Live model for eligible paid ChatGPT plans, while GPT-Live-1 mini serves free users. Live is available through supported iOS, Android and web experiences, although rollout timing and usage limits depend on plan, workspace, region and app version. A single Live conversation can last up to two hours.[4]
Advanced Voice and Standard Voice remain selectable where available. Advanced is still required for supported video and screen-sharing sessions, which Live did not support at launch. Live also does not initially support custom GPTs, connected apps or plugins.[4]
OpenAI says audio clips from Live and Advanced Voice conversations are retained for 30 days alongside the transcript in chat history. Deleting the chat triggers deletion of associated clips within 30 days, subject to security, safety, legal and previously disassociated training-data exceptions. Audio clips are not used for training unless an eligible personal-workspace user opts in, although transcripts and other conversation files may be used when the broader model-improvement setting is enabled. Standard Voice audio is deleted after transcription unless the user has chosen to share it.[5]
For safety, OpenAI says it checks inputs and outputs as conversations unfold and can redirect a response, play a warning, display support resources or end a call. Its system card reports generally equal or better results than Advanced Voice on selected voice-native safety tests, but those evaluations were deliberately adversarial and were not weighted to represent failure rates in normal use. GPT-Live-1 scored below Advanced Voice on one emotional-reliance test, a difference OpenAI said was not statistically significant.[6] The remaining product boundary is developer access. OpenAI’s August 3 engineering post says the architecture will underpin an upcoming GPT-Live API. As of August 4, the company’s API catalog instead lists models including GPT-Realtime-2.1 for production speech-to-speech applications. GPT-Live is therefore production technology inside ChatGPT, but not yet a generally available API model that outside developers can deploy directly.[2][7]
Companies mentioned #
Further sources #
[[1] OpenAI, “Introducing GPT-Live,” July 8, 2026. ↗](https://openai.com/index/introducing-gpt-live/)
[[2] OpenAI, “How we built a realtime system for responsive voice AI in six months,”… ↗](https://openai.com/index/continuous-voice-interaction-with-gpt-live/)
[[3] Agora Media Lab, “OpenAI Didn’t Publish GPT-Live’s Latency. So We Measured It,”… ↗](https://www.agora.io/en/blog/openai-didnt-publish-gpt-lives-latency-so-we-measured-it/)
[[4] OpenAI Help Center, “ChatGPT Voice,” accessed August 4, 2026. ↗](https://help.openai.com/en/articles/20001274/)
[[5] OpenAI Help Center, “ChatGPT Voice — Data controls,” accessed August 4, 2026. ↗](https://help.openai.com/en/articles/20001274/)
[[6] OpenAI, “GPT-Live System Card,” July 8, 2026. ↗](https://deploymentsafety.openai.com/gpt-live)+1 more
The stories that matter, in one email. Free — unsubscribe anytime.