GPT-Live

GPT-Live is a family of full-duplex voice models developed by 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 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] On September 10, 2026, OpenAI released GPT-Live-1 to developers as the API model gpt-live-1, priced at $0.05 per minute for what the company calls the front-end voice layer.^[10][11]

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 Instant. The Medium and High settings used GPT-5.5 Thinking with medium or high reasoning effort.^[1] In the API the split is explicit: the developer picks the backend model and agent harness separately from the voice model, and pays for each separately.^[10][12]

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 and Standard Voice remained available as separate options. OpenAI's help page said at launch that Live was not available in ChatGPT Business, Enterprise, or Edu workspaces, and that users in those workspaces could keep the Voice options they already had. Workspace support arrived in late July 2026: a July 23, 2026 capture of the page has no workspace section, while an August 2, 2026 capture describes a two-week early access period in which Enterprise, Edu, and Healthcare owners had to enable both Voice and Early Model Access before members could use Live. As of September 11, 2026 the page says Voice is available in eligible Business, Enterprise, Edu, and Healthcare workspaces subject to workspace settings, that Live is the default Voice experience once a workspace owner enables Voice, and that Early Model Access is not required.^[3][4][30]

The three ChatGPT Voice options use different interaction designs:

GPT-Live is not the same product as the developer-facing OpenAI Realtime API, and the two reached developers at different times. OpenAI said at the consumer launch that it planned to add GPT-Live-1 to the API and opened a notification form, and an August 3, 2026 engineering post still described the GPT-Live API as upcoming.^[1][5][8] Developer access arrived on September 10, 2026 with a post titled "Build more natural voice experiences with GPT-Live-1 in the API," tagged Product Release.^[10] OpenAI posted a 77-second video with the same news to its YouTube channel the following day.^[19] The API model catalog and pricing page list gpt-live-1 alongside, not in place of, the GPT-Realtime family, and OpenAI's documentation frames the two as different architectures rather than successive versions of one.^[11][13][15]

API release

OpenAI published the API release on September 10, 2026. The announcement describes GPT-Live-1 as having been introduced first in ChatGPT, and presents the API version as the same full-duplex model with new controls for developers. Its framing of the alternative is a chained pipeline: "Traditional voice agents stitch together speech-to-text, a reasoning model, and text-to-speech. Each handoff adds latency and creates more opportunities to lose timing, context, or the natural rhythm of a conversation."^[10]

The strengths OpenAI listed for the API release were interruption handling through a single model that reasons over incoming and outgoing audio together; delegation of reasoning and tool calls to a backend text model such as GPT-6 Astra or a third-party model; control of an agent's tone, pace, and conversational style through the system prompt; handling of background noise and silence without interrupting or narrating; context retention across long sessions; and telephony support for phone-call agents. OpenAI also said the model natively provides automatic speech recognition transcripts and response text, offers alphanumeric understanding, supports keyword biasing, and supports turn detection natively even though it is not a turn-based model.^[10]

Model and session specifications

OpenAI's model page and session documentation give the following:

Rate limits are counted in concurrent sessions rather than tokens per minute, and the free usage tier is not supported:^[11]

Model, instructions, startup history, audio configuration, voice, delegation mode, and storage are all fixed when the session is created. Only Responses-delegation settings can be updated in a running session; changing the model, the voice, or the delegation mode requires a new session.^[13]

Pricing

GPT-Live-1 is billed by session duration rather than by token. The rate is $0.05 per minute, billed per second and not rounded up to a whole minute, and it covers the voice layer only: backend model calls and tool use are billed at their own rates.^[11][15][16] A POST /v1/live/sessions request that creates a WebRTC session bills 15 seconds of voice duration during initialization, credited against the running session's charges rather than added to them. Billable time covers the whole active session, including stretches where nobody is speaking or the backend is working, so muting a microphone does not stop the meter.^[16] OpenAI's own worked example puts a 90-second voice session at $0.075 and, with $0.02 of backend model and tool cost, a total of $0.095 for the conversation.^[16]

That is a different billing shape from the Realtime line it is compared against. As of the September 2026 pricing page, gpt-realtime-2.1 is billed per token at $32.00 per million audio input tokens and $64.00 per million audio output tokens, plus separate text and image rates, so the two cannot be compared on a single published number.^[15]

Delegation

The API exposes two delegation modes, chosen at session creation:^[14]

In either mode the application still owns permissions, confirmations, business records, and durable task state, and interrupting the assistant's speech does not by itself cancel backend work.^[12][14] Applications feed the live model through three append events, each a plain string capped at 500 tokens: session.thinking.append for background context the model should not say aloud, session.commentary.append for a result the model should speak and is trained to paraphrase, and session.instructions.append for session-wide steering.^[14] The announcement's code sample uses the Codex SDK as a client-delegation backend, starting a Codex thread against a repository and returning its final response through session.commentary.append.^[10]

Connections, sessions, and voices

Browser applications connect over WebRTC, server-owned audio streams over WebSockets, and phone systems through telephony and SIP paths; a separate sideband control connection lets a trusted server observe and steer an existing session. OpenAI documents partner integrations for LiveKit, Twilio, Telnyx, and Daily/Pipecat, and warns that an existing Realtime integration is not automatically compatible with GPT-Live.^[12][18]

Context is managed without a configuration parameter. GPT-Live summarizes older history in the background, and once context usage passes 90 percent it starts a replacement voice engine inside the same session that receives the original instructions and up to 8,192 tokens of conversation history. Session recordings are off by default; enabling store makes a completed session available for download or forking, and recordings expire after 30 days.^[13]

The API adds twelve named voices beyond the default, marin, and OpenAI describes the regional labels as a speaking style rather than a guarantee of accent fidelity:^[10][13]

OpenAI said it would keep expanding voice and language availability over the following months.^[10] Custom voices remain limited to eligible customers who contact sales. Creating one requires a consent recording in which the voice actor reads a fixed phrase, plus a reference sample from the same speaker; an organization may hold at most 20 custom voices. For GPT-Live specifically, custom voices are supported with English accents, the accent also has to be requested in the session instructions, and the custom voice is passed as an object rather than the plain string used for named voices.^[10][17]

Another route OpenAI named is OpenAI Presence, an enterprise product that uses GPT-Live-1 to power real-time voice interactions.^[10]

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] The same caveat carries into the API benchmarks, where OpenAI names the backend model used for several of the headline scores.^[10]

OpenAI did not publish a single end-to-end latency number at the ChatGPT 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 mid-conversation. The API documentation describes the developer-visible form of this behavior: above 90 percent context usage, a replacement voice engine starts inside the same session with the original instructions and a bounded slice of recent history.^[8][13] 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 would underpin the then-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 desktop voice experience on macOS and Windows, in which the voice interface can start tasks, check progress, ask about running agents, and coordinate several agents through one conversation. It was called Voice in Work and Codex in the August 2, 2026 version of the page and is listed as Voice in Desktop as of September 11, 2026; the API announcement points at the same pairing with ChatGPT Work and Codex as an example of delegation.^[3][10][30]

OpenAI provides nine preset ChatGPT voices for Live, and a separate, larger set in the API. The system is designed for conversation rather than voice cloning, and its safeguards are intended to prevent imitation of a real person's voice; custom voices in the API require a recorded consent statement from the speaker.^[1][2][17] OpenAI has not released the model weights, parameter counts, or model-level architectural details beyond the full-duplex and delegation design. The API release did publish figures that the ChatGPT launch had left out, including the July 31, 2025 knowledge cutoff, the 128,000-token default context window, the supported modalities, and the concurrent-session rate limits.^[11][13]

Access and usage limits

Consumer access is bundled into ChatGPT subscriptions and governed by rolling usage limits rather than per-token billing. OpenAI revises the help page that publishes these limits frequently, and the page itself says limits may change, so any figure taken from it is a snapshot of one day. The version archived on July 23, 2026 listed only Free, Go and Plus, and the two Pro tiers, split each paid allowance between Instant and Medium or High intelligence, and said a single Live conversation could last up to 2 hours. None of those three features survive: the plan list is longer, the intelligence split is gone, and the page now refers to a maximum session length without publishing a figure for it. Read on September 11, 2026, when it was marked as updated within the previous hour, the page gave the following allowances for Voice in Chat, measured over a rolling 24-hour period:^[3][30]

The desktop experience is listed separately as Voice in Desktop, with its availability and allowances moved to another help page rather than stated here.^[3] These plan entitlements are product pricing rather than a price for running GPT-Live as a standalone model. Developer access is priced separately, at $0.05 per minute of voice session plus backend costs.^[11][15]

Evaluation

ChatGPT launch evaluations

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, 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]

API launch benchmarks

The API announcement carried seven charts comparing GPT-Live-1 with gpt-realtime-2.1 and gpt-realtime-2. Four of them carry a footnote naming the backend model that produced the GPT-Live row, which matters because a score achieved by GPT-Live-1 paired with a reasoning model is not a score for the voice model on its own. The figures below are OpenAI's own, taken from the chart data on the announcement page:^[10]

OpenAI's captions define each chart. Tau3 (Voice) Intelligence covers spoken customer-service tasks in airline, retail, and telecom domains, with the headline weighting each domain equally. Tau Banking (Voice) Knowledge is the fraction of 97 banking_knowledge tasks completed successfully. The Artificial Analysis chart measures pause handling, turn taking, interruptions, and backchannels. Full Duplex Bench v1.5 Interactivity tests reactions to background speech, speech to another person, listener backchannels, and interruptions. Full Duplex Bench v1 Turn-taking latency measures how quickly the agent starts its reply after the user finishes a turn. The two v3 charts test tool use and spoken answers when requests contain natural pauses, hesitations, and self-corrections.^[10]

The two summary claims in the announcement text are OpenAI's: that GPT-Live-1 "improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1," and that, paired with GPT-6 Astra at medium reasoning effort, "it also ranks #1 on Tau3, which measures frontier voice-agent intelligence on end-to-end tasks."^[10]

What the cited benchmarks are

Full Duplex Bench is an academic benchmark suite from Guan-Ting Lin and collaborators, with code and data on GitHub. Version 1 (arXiv 2503.04721, submitted March 2025) evaluates pause handling, backchanneling, smooth turn-taking, and user interruption with automatic metrics. Version 1.5 (arXiv 2507.23159, July 2025) adds four overlap scenarios: user interruption, user backchannel, talking to others, and background speech, which is the same list OpenAI's v1.5 caption describes. Version 2 (arXiv 2510.07838, October 2025, revised April 2026) is a streaming framework paired with an automated examiner that enforces staged goals across multiple turns at two pacing settings, over four task families: daily, correction, entity tracking, and safety.^[22] Version 3 (arXiv 2604.04847, April 2026) pairs real human recordings annotated for five disfluency types with scenarios that require chained API calls across four task domains; its code and data were released in May 2026.^{[20][21][23][24]}

The v3 paper's own results evaluated six configurations, including GPT-Realtime, Gemini Live 2.5, Gemini Live 3.1, Grok, Ultravox v0.7, and a cascaded Whisper to GPT-4o to text-to-speech pipeline. It reported GPT-Realtime leading on Pass@1 at 0.600, and identified self-correction handling and multi-step reasoning on hard scenarios as the most consistent failure modes across systems.^[23]

Tau3 refers to the third generation of Sierra's tau-bench line of tool-agent-user benchmarks, announced on March 18, 2026. It extends the earlier text-only benchmark with a knowledge-retrieval domain called banking_knowledge and with full-duplex voice evaluation against real-time audio providers; its domains are airline, retail, telecom, banking_knowledge, and a mock domain. The voice component is documented in the tau-Voice paper (arXiv 2603.13686) and the knowledge component in tau-Knowledge (arXiv 2603.04370).^[25][26][27] Comparability across versions is a live issue in that benchmark: Sierra's July 2026 v1.0.1 release fixed errors in banking_knowledge tasks and states that results produced before 1.0.1 are not comparable with 1.0.1 and later. OpenAI's chart labels and captions do not name a tau-bench release version.^[10][25]

Artificial Analysis Conversational Dynamics is a benchmark that Artificial Analysis implements itself from a subset of Full Duplex Bench v1 and v1.5. Its methodology page lists four component metrics: pause handling and turn taking from v1, and user interruption handling and backchannel handling from v1.5.^[28]

Independent measurements

Artificial Analysis publishes its own speech-to-speech results, and versions its datasets, so a figure taken from it needs a version as well as a date. Its methodology page records Big Bench Audio at v1.2 since May 2026 and the Artificial Analysis Speech to Speech Index at v2.0 since August 2026, the release that removed Conversational Dynamics from the index in favor of a task success rate. As read on September 11, 2026 the leaderboard listed two GPT-Live-1 entries, distinguished by the backend model and reasoning effort behind them, under model slugs containing diamond-alpha:^[28][29]

Two things follow from putting the two sources side by side. First, the Conversational Dynamics chart in OpenAI's announcement carries the numbers Artificial Analysis itself publishes. The underlying values for the two GPT-Realtime rows are identical to Artificial Analysis's figures for GPT-Realtime-2.1 High and GPT-Realtime-2 (High) to every published digit, and OpenAI's 97.3 percent for GPT-Live-1 is Artificial Analysis's 97.27 percent figure for GPT-Live-1 (Sol, low), rounded. OpenAI's chart does not name a backend for that row; the Astra-medium configuration scores 94.9 percent on the same measure.^[10][29]

Second, the two sources differ substantially on voice-agent task success for the same named configuration. OpenAI reports 86.2 percent Pass@1 on Tau3 (Voice) Intelligence for GPT-Live-1 with an Astra medium backend, while Artificial Analysis reports 67.9 percent on tau-Voice for the entry it labels GPT-Live-1 (Astra, medium), over the same three domains with equal weighting. The two are separate implementations: Artificial Analysis runs its own harness with its own simulated-customer voice personas, and its methodology states that each tau-Voice score is the mean of three independent trials, while its own data for both GPT-Live-1 entries records a single trial.^[10][28][29]

Customer reports

OpenAI's announcement carried attributed quotes from four named companies plus a fifth quote whose speaker is named but whose company appears only as a logo on the testimonial card:^[10]

These are customer statements reproduced by the vendor, not independent measurements, and none of them is accompanied by a published methodology. OpenAI repeated the Speak figure in its own body text as evidence that interruption handling was "already delivering business impact."^[10] The page also carried recorded demonstrations from Yelp, Cognition, Picsart, HeyGen, and OpenAI Presence, and by the September 11 capture it had added a time-limited in-browser demo session.^[10]

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, 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]

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 0.92. 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]

Those are OpenAI's corrected figures, not the ones it published at launch. On August 4, 2026 the company republished the card with a change log saying it had found "a configuration mismatch in the safety evaluation setup" and that the original results had been generated "using a backend configuration that did not match the final model release". It re-ran the full evaluation with the correct configuration, updated the results in the body of the card, and added an appendix setting the original and corrected values side by side. Measured against the values it had previously reported, the corrected production scores are statistically significant regressions on illicit behavior for GPT-Live-1 (0.97 to 0.95) and GPT-Live-1 mini (0.94 to 0.91); on the synthetic set, GPT-Live-1 regresses on sexual content (0.97 to 0.95) and gore (0.97 to 0.93), and GPT-Live-1 mini on illicit behavior (0.97 to 0.95) and gore (0.96 to 0.90). OpenAI said the overall safety conclusions were unchanged and that none of the regressions fell below its safety launch standards. The July version of the card had put the mini model's production sexual-content score at 0.95 rather than 0.92 and had added, of that regression and the emotional-reliance one, "Note that neither of these are statistically significant"; the corrected card drops that sentence, and its only remaining use of the phrase is in the appendix comparing the original and corrected runs.^[2][31]

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 twelve API voices are labelled English and Brazilian Portuguese only, and OpenAI describes their regional labels as a speaking style rather than a guarantee of accent fidelity.^[13]

Benchmark interpretation is complicated by the delegation design. Several of OpenAI's headline numbers were produced with a named reasoning backend, so they measure a voice model plus a text model rather than the voice model alone, and the two named backends differ between charts. Where an independent evaluator has published comparable figures, they do not always agree with OpenAI's: Artificial Analysis reports a markedly lower tau-Voice score for the same named GPT-Live-1 and backend configuration.^[10][29] The tau-bench line has also changed its grading within the past year, and OpenAI's charts do not state which release they used.^[25]

The API release narrowed the reproducibility gap that existed at the ChatGPT launch, since researchers can now create sessions directly, choose the backend, and hold settings fixed. Other disclosures did not follow: the model page publishes modalities, the knowledge cutoff, endpoints, and rate limits, but no parameter count, training-compute figure, or architectural specification.^[10][11]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12OpenAI. "Introducing GPT-Live." July 8, 2026. openai.com/...introducing-gpt-live
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10OpenAI. "GPT-Live System Card." Published July 8, 2026; safety evaluation results corrected August 4, 2026. Accessed September 11, 2026. deploymentsafety.openai.com/gpt-live
3. ^1 ^2 ^3 ^4 ^5 ^6OpenAI Help Center. "ChatGPT Voice." Read September 11, 2026, when the page was marked as updated within the previous hour. help.openai.com/...20001274-gpt-live-faq
4. ^OpenAI Help Center. "ChatGPT Release Notes." July 8, 2026 entry. help.openai.com/...6825453-chatgpt-release-notes
5. ^OpenAI. "GPT-Live-1 in the API." Accessed July 24, 2026. openai.com/...gpt-live-1-in-the-api
6. ^Ivan Mehta. "OpenAI releases new voice models for more natural live conversations." TechCrunch, July 8, 2026. techcrunch.com/...-more-natural-live-conversations
7. ^1 ^2Hermes Frangoudis. "OpenAI Didn't Publish GPT-Live's Latency. So We Measured It." Agora Media Lab, July 10, 2026. prod.agora.io/...t-lives-latency-so-we-measured-it
8. ^1 ^2 ^3 ^4 ^5 ^6Justin Uberti and Zahan Malkani. "How we built a realtime system for responsive voice AI in six months." OpenAI, August 3, 2026. openai.com/...uous-voice-interaction-with-gpt-live
9. ^1 ^2OpenAI (@OpenAI). "GPT-Live can listen while it speaks." X post, August 3, 2026. x.com/...2084378415818579975
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23OpenAI. "Build more natural voice experiences with GPT-Live-1 in the API." September 10, 2026. openai.com/...introducing-gpt-live-1-in-the-api (read via the Internet Archive captures of September 10 and September 11, 2026; chart values and footnotes read from the page's embedded chart data)
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7OpenAI. "GPT-Live 1" model page. Accessed September 11, 2026. developers.openai.com/...gpt-live-1
12. ^1 ^2 ^3OpenAI. "Getting started with GPT-Live." Accessed September 11, 2026. developers.openai.com/...live
13. ^1 ^2 ^3 ^4 ^5 ^6 ^7OpenAI. "Managing GPT-Live sessions." Accessed September 11, 2026. developers.openai.com/...live-conversations
14. ^1 ^2 ^3OpenAI. "Delegation and tools in GPT-Live." Accessed September 11, 2026. developers.openai.com/...live-delegation
15. ^1 ^2 ^3 ^4OpenAI. "Pricing." Accessed September 11, 2026. developers.openai.com/...pricing
16. ^1 ^2 ^3OpenAI. "Cost optimization" (voice latency and cost). Accessed September 11, 2026. developers.openai.com/...voice-latency-cost
17. ^1 ^2OpenAI. "Custom voices." Accessed September 11, 2026. developers.openai.com/...custom-voices
18. ^OpenAI. "GPT-Live partner integrations." Accessed September 11, 2026. developers.openai.com/...live-partner-integrations
19. ^OpenAI. "GPT-Live-1 is now in the API." YouTube, September 11, 2026, 77 seconds. youtube.com/watch
20. ^Guan-Ting Lin, Jiachen Lian, Tingle Li, Qirui Wang, Gopala Anumanchipalli, Alexander H. Liu, Hung-yi Lee. "Full-Duplex-Bench: A Benchmark to Evaluate Full-duplex Spoken Dialogue Models on Turn-taking Capabilities." arXiv:2503.04721, March 6, 2025. arxiv.org/...2503.04721
21. ^Guan-Ting Lin, Shih-Yun Shan Kuan, Qirui Wang, Jiachen Lian, Tingle Li, Shinji Watanabe, Hung-yi Lee. "Full-Duplex-Bench v1.5: Evaluating Overlap Handling for Full-Duplex Speech Models." arXiv:2507.23159, July 30, 2025. arxiv.org/...2507.23159
22. ^Guan-Ting Lin, Shih-Yun Shan Kuan, Jiatong Shi, Kai-Wei Chang, Siddhant Arora, Shinji Watanabe, Hung-yi Lee. "Full-Duplex-Bench-v2: A Multi-Turn Evaluation Framework for Duplex Dialogue Systems with an Automated Examiner." arXiv:2510.07838, October 9, 2025, revised April 26, 2026. arxiv.org/...2510.07838
23. ^1 ^2Guan-Ting Lin, Chen Chen, Zhehuai Chen, Hung-yi Lee. "Full-Duplex-Bench-v3: Benchmarking Tool Use for Full-Duplex Voice Agents Under Real-World Disfluency." arXiv:2604.04847, April 6, 2026. arxiv.org/...2604.04847
24. ^Full-Duplex-Bench repository. Accessed September 11, 2026. github.com/...Full-Duplex-Bench
25. ^1 ^2 ^3Sierra Research. tau2-bench repository README, describing tau3-bench. Accessed September 11, 2026. github.com/...tau2-bench
26. ^Sierra. "tau3-bench: Voice." March 18, 2026. sierra.ai/...tau-3-bench
27. ^Soham Ray, Keshav Dhandhania, Victor Barres, Karthik Narasimhan. "tau-Voice: Benchmarking Full-Duplex Voice Agents on Real-World Domains." arXiv:2603.13686, 2026. arxiv.org/...2603.13686
28. ^1 ^2 ^3Artificial Analysis. "Speech to Speech Benchmarking Methodology." Accessed September 11, 2026. artificialanalysis.ai/...ch-to-speech-benchmarking
29. ^1 ^2 ^3 ^4Artificial Analysis. "Speech to Speech" leaderboard. Accessed September 11, 2026. artificialanalysis.ai/speech-to-speech
30. ^1 ^2 ^3OpenAI Help Center. "ChatGPT Voice." Internet Archive captures of July 12, 2026, July 23, 2026, and August 2, 2026. web.archive.org/...20001274-chatgpt-voice ; web.archive.org/...20001274-chatgpt-voice ; web.archive.org/...20001274-chatgpt-voice
31. ^OpenAI. "GPT-Live System Card." Internet Archive capture of July 9, 2026, the version published before the August 4, 2026 correction. web.archive.org/...gpt-live