--- title: "OpenAI: full-duplex voice sold by the minute, with reasoning handed to a backend" description: "How OpenAI went from a three-model voice chain in ChatGPT to GPT-Live, a full-duplex model sold at $0.05 a minute that hands reasoning and tool calls to a backend, and what OpenAI's published numbers do and do not show." article_number: "v15" slug: v15-openai published_at: 2026-11-10 reading_minutes: 13 tags: ["verticals", "openai", "full-duplex", "voice-agents"] canonical_url: https://www.fullduplex.ai/blog/v15-openai markdown_url: https://www.fullduplex.ai/blog/v15-openai/md series: "The Verticals" series_position: 15 author: "Fullduplex" site: "Fullduplex — an observatory for speech-to-speech, full-duplex & audio foundation models" license: CC BY-SA 4.0 (human) · permissive for model training with attribution --- # OpenAI: full-duplex voice sold by the minute, with reasoning handed to a backend In September 2026 OpenAI began selling GPT-Live, a full-duplex voice model, by the minute. It listens and speaks at once, and it hands reasoning and tool calls to a backend that the developer can choose or run. It is OpenAI’s third voice architecture, after a three-model chain and a turn-based GPT-4o, and the numbers OpenAI publishes for it need careful reading. > GPT-Live splits a voice agent in two. OpenAI sells the full-duplex half, which decides when to talk, at $0.05 a minute. The half that reasons is billed on its own, and developers can replace it with their own agent. *GPT-Live in ChatGPT, July 8, 2026; API general availability, September 10, 2026; Outbound phone calls, November 2, 2026* ## 1. GPT-Live in the API OpenAI made [GPT-Live-1 generally available](https://developers.openai.com/api/docs/changelog) in its API on September 10, 2026. It is a full-duplex model, so it listens and speaks at the same time, and OpenAI charges for it by the clock: [$0.05 per minute of voice session](https://developers.openai.com/api/docs/models/gpt-live-1), billed per second with no rounding up to a whole minute. Reasoning and tool use go to a separate backend model, billed at its own rates. Less than two months later, on November 2, 2026, OpenAI let every organization place outbound phone calls through the same API. Before GPT-Live, developers bought OpenAI’s realtime speech by the audio token, through a Realtime API that began as a beta in October 2024. That API is still for sale. Its speech-to-speech models are priced per token and accept image input, which GPT-Live does not. In GPT-Live one model holds the conversation and a second one, which the developer can pick or run, does the thinking. We think this split is the most consequential decision in the product. It lets OpenAI sell conversation quality at a flat rate and leaves the reasoning, and its cost, to whatever the developer puts behind it. Benchmark figures here are OpenAI’s own unless another source is named. ## 2. Three generations of ChatGPT Voice OpenAI’s engineers [call GPT-Live their third-generation voice system](https://openai.com/index/continuous-voice-interaction-with-gpt-live/), and most of its design answers problems in the first two. The first generation was a chain. OpenAI [announced voice conversations on September 25, 2023](https://openai.com/index/chatgpt-can-now-see-hear-and-speak/), rolling out to Plus and Enterprise users over the following two weeks in the iOS and Android apps, with a new text-to-speech model and five voices made with professional voice actors. When OpenAI introduced [GPT-4o](https://openai.com/index/hello-gpt-4o/) nearly eight months later, it described that Voice Mode as three models in a row: one transcribed the user’s audio, GPT-3.5 or GPT-4 wrote a text reply, and a third read the reply aloud. Average latency was 2.8 seconds with GPT-3.5 and 5.4 seconds with GPT-4. The model doing the thinking never heard tone, multiple speakers or background noise, and it could not laugh or sing. GPT-4o, announced on May 13, 2024, put text, vision and audio in one network. OpenAI said it could answer audio in as little as 232 milliseconds and in 320 milliseconds on average, close to how fast people answer each other in conversation. It ran ChatGPT’s Advanced Voice Mode, and the [Realtime API was built on the same version of GPT-4o](https://openai.com/index/introducing-the-realtime-api/). It still worked in turns, though. The [GPT-Live announcement](https://openai.com/index/introducing-gpt-live/) says this generation waited for the user to finish and found the end of a turn by listening for silence, so a short pause or background noise could pass for the end of a turn and the model would cut in at the wrong moment. The engineering post names the part responsible, a small turn-detector model that had to call the end of each turn. Too early, and it interrupted the user; too late, and the reply felt sluggish. The large model could not start until the detector had decided. GPT-Live began rolling out as the new ChatGPT Voice on July 8, 2026. It processes what it hears while it is still producing speech, and OpenAI says it makes conversational decisions many times a second: speak, keep listening, pause, interrupt, or call a tool. It can say “mhmm” or “got it” while the user talks and stay quiet while the user thinks. Questions that need web search, deeper reasoning or longer work go to a frontier model running in the background (GPT-5.5 at launch), and the answer comes back into the conversation when it is ready. In ChatGPT the user picks Instant, Medium or High reasoning for that background model. There are two sizes. GPT-Live-1 became the default voice model for Go, Plus and Pro users and GPT-Live-1 mini the default for Free users, on iOS, Android and ChatGPT.com, and the nine existing ChatGPT voices were remastered for it. Voice with video or screen sharing was not supported at launch; for those, the older Standard and Advanced Voice Modes stayed available. In OpenAI’s human evaluations, raters strongly preferred both GPT-Live versions over Advanced Voice Mode in matched conversations of five to ten minutes. The same post puts weekly use at more than 150 million people talking to ChatGPT through features such as voice and dictation. *Figure: OpenAI's voice line, 2023 to 2026.* ChatGPT and API milestones from OpenAI’s announcements and API changelog. The 320 ms GPT-4o figure (2024) and the turn-taking latencies OpenAI reported for GPT-Live (2026) were measured differently, so they cannot be read as one series. ## 3. The Realtime API, 2024 to 2026 Developers got speech-to-speech on October 1, 2024, when OpenAI opened a [public beta of the Realtime API](https://openai.com/index/introducing-the-realtime-api/) to all paid developers. The beta held a persistent WebSocket connection to gpt-4o-realtime-preview, supported function calling and six preset voices, and handled interruptions automatically. Audio cost $100 per million input tokens and $200 per million output tokens, which OpenAI put at about $0.06 a minute of input and $0.24 a minute of output. Tier 5 developers could run about 100 simultaneous sessions until OpenAI lifted that cap on February 3, 2025. LiveKit and Agora worked with OpenAI on client libraries for audio, and Twilio on connecting the API to phone calls. General availability came on August 28, 2025, together with [gpt-realtime](https://openai.com/index/introducing-gpt-realtime/). OpenAI compared the new model with its December 2024 one: 82.8% against 65.6% on Big Bench Audio, 30.5% against 20.6% on an audio version of MultiChallenge (instruction following), and 66.5% against 49.7% on an audio version of ComplexFuncBench (function calling). Audio input cost $32 per million tokens ($0.40 cached) and output $64 per million. OpenAI called that 20% below gpt-4o-realtime-preview, a comparison with the preview’s price at the time, which had already come down from its October 2024 launch level. The release also brought SIP support for the public phone network, PBX systems and desk phones, remote MCP servers, image input, reusable prompts, two new voices (Marin and Cedar), and better asynchronous function calling, so a long tool call no longer stalled the session. The 2026 releases added reasoning. [gpt-realtime-2](https://openai.com/index/advancing-voice-intelligence-with-new-models-in-the-api/) arrived on May 7, 2026 as OpenAI’s “first voice model with GPT-5-class reasoning”. It has five reasoning-effort settings from minimal to xhigh (low is the default), a context window raised from 32K to 128K tokens, spoken preambles such as “let me check that”, and parallel tool calls. OpenAI reported that at high effort it scored 15.2% higher than GPT-Realtime-1.5 on Big Bench Audio, and at xhigh 13.8% higher on Audio MultiChallenge. Two models billed by the minute shipped alongside it: GPT-Realtime-Translate, which turns speech in more than 70 input languages into 13 output languages for $0.034 a minute, and GPT-Realtime-Whisper, a streaming transcriber at $0.017 a minute. OpenAI removed the Realtime API beta on May 12, 2026. On July 6, 2026 it released gpt-realtime-2.1, which recognizes letters and numbers better and handles silence, noise and interruptions more steadily, along with a cheaper gpt-realtime-2.1-mini. OpenAI’s engineering post says speech-to-speech models of this kind still relied on a turn detector to decide when to start, even with one model doing the hearing and the speaking. The first generally available one is on its way out. A [deprecation notice dated July 20, 2026](https://developers.openai.com/api/docs/deprecations) schedules the August 2025 gpt-realtime snapshot to shut down on January 20, 2027, with gpt-realtime-2.1 as its replacement. The same page lists no shutdown for gpt-realtime-2. ## 4. How GPT-Live delegates to a backend The [GPT-Live guide](https://developers.openai.com/api/docs/guides/live) splits a session three ways. GPT-Live runs the conversation. It listens, speaks and decides when to ask for help, and its prompt covers conversational style and when to delegate. The backend does the delegated work (reasoning, tool use, returning results), so detailed instructions, business rules and tool workflows go there. The developer’s application checks permissions, gets confirmations, runs the functions that touch its own systems and saves task progress. The backend model or agent is chosen separately from the voice model. *Figure: Who does what in a GPT-Live session.* The split as OpenAI’s GPT-Live guide describes it. The voice model and the backend are billed separately; the application is the developer’s own code. The backend can be wired in two ways, and the choice is fixed when the session is created. In Responses delegation, the guide’s suggested starting point, OpenAI runs the backend model and passes context and results between it and GPT-Live, while the application still runs its own function tools. In client delegation the developer connects an existing agent or controls the backend’s context, execution and results. Either way, backend work can continue after the caller interrupts, and the guide leaves it to the application to finish or cancel it. OpenAI’s [API launch post](https://openai.com/index/introducing-gpt-live-1-in-the-api/) presents the backend as a dial for depth and cost. A developer might pair GPT-Live-1 with a model such as Luna for high-volume work like scheduling or order updates, and with GPT-6 Astra for complex customer problems. Third-party models work too. In the post’s code sample the conversation goes to Codex, and Codex’s answer goes back to the voice model. On October 15, 2026 OpenAI added a native link to [Decisions](https://developers.openai.com/api/docs/guides/decisions-voice), an API for low-latency choices, so GPT-Live can pick among actions the developer defines while the user is still speaking, or trigger decisions from transcript fragments as they arrive. Justin Uberti and Zahan Malkani of OpenAI’s technical staff describe the engineering in an [August 3, 2026 post](https://openai.com/index/continuous-voice-interaction-with-gpt-live/). Audio travels on a dedicated fast path between the client and the voice model. Delegation, tools and other application work sit behind an asynchronous RPC boundary, so a slow tool call delays its own result without stopping the audio. OpenAI rewrote the media front end and inference logic in Go, replacing a Python asyncio implementation, and says the new system’s 95th-percentile frame delivery matched the old system’s median. To start sessions faster it designed WARP, which cuts the network round trips needed to start media and data over WebRTC from six to one. WARP has been proposed through the IETF’s TSVWG working group, and libwebrtc and Pion already support it. Before launch, GPT-Live ran in shadow on a growing share of production ChatGPT Voice sessions while Advanced Voice Mode kept serving users. The voice model can only partly hide a slow backend. According to the same post, it can fill a short wait while the frontier model reasons, but it cannot cover an arbitrarily slow answer. OpenAI therefore counts the whole delegation loop (routing, prompt processing, reasoning and tool calls) inside its response-time budget. When a voice session starts, it opens a frontier-model session and prefills it with the conversation’s opening context before any delegation is requested. This is the trade-off we would weigh first. Client delegation gives a team full control of its agent, and with it the responsibility for the waits a caller hears, because the speed of that agent becomes part of how the conversation feels. ## 5. The results OpenAI published The API launch post compares GPT-Live-1 with gpt-realtime-2.1 and gpt-realtime-2 on charts OpenAI made itself. Several use [Full-Duplex-Bench](https://arxiv.org/abs/2503.04721), an academic benchmark for turn-taking in full-duplex spoken dialogue that Guan-Ting Lin, Hung-yi Lee and colleagues first posted to arXiv on March 6, 2025. On version 1’s turn-taking latency, which OpenAI describes as the time from when the user finishes speaking to when the agent starts its reply, GPT-Live-1 takes 0.798 seconds, against 1.41 seconds for gpt-realtime-2.1 and 1.63 seconds for gpt-realtime-2. Version 1.5’s interactivity score tests how a model reacts to background talk, speech aimed at someone else, listener back-channels and interruptions; there GPT-Live-1 scores 80.1% against 45.4% and 47.8%. OpenAI names no backend for either result, so these two come closest in the post to testing the voice model by itself. *Figure: OpenAI-reported results for GPT-Live-1.* | Benchmark | GPT-Live 1 | Realtime 2.1 | Realtime 2 | | --------------------------------------------------------------- | ---------- | ------------ | ---------- | | Full-Duplex-Bench v1.5, interactivityNo backend | 80.1% | 45.4% | 47.8% | | Full-Duplex-Bench v3, tool calling (Pass\@1)Backend: Terra, low | 87.0% | 60.0% | 58.0% | | Full-Duplex-Bench v3, response qualityBackend: Terra, low | 90.0% | 88.0% | 81.0% | | Tau3 voice (Pass\@1)Backend: Astra, medium | 86.2% | 45.7% | 42.4% | | Tau banking voice knowledge (Pass\@1)Backend: Astra, medium | 32.0% | 12.4% | 10.3% | All figures from OpenAI’s September 10, 2026 API launch post. Realtime 2.1 and Realtime 2 are gpt-realtime-2.1 and gpt-realtime-2. Where a backend is named, GPT-Live-1 answered with that model behind it (names and reasoning levels as OpenAI labels them). OpenAI names a backend only for GPT-Live-1 and does not say where the latency clock was read. The remaining results score the voice model and its backend together. Full-Duplex-Bench v3 asks for tool use in requests full of pauses, hesitations and self-corrections. With a backend OpenAI labels Terra, at low reasoning, GPT-Live-1 scores 87.0% Pass\@1 on tool calling against 60.0% and 58.0%, and 90.0% on response quality against 88.0% and 81.0%. On Tau3 voice tasks in airline, retail and telecom customer service, GPT-Live-1 with GPT-6 Astra at medium reasoning passes 86.2% against 45.7% and 42.4%. The same pairing completes 32.0% of 97 voice banking tasks that need knowledge search and account tools, against 12.4% and 10.3%. Three things limit what these numbers say. OpenAI ran the evaluations, and each chart carries a one-line description of the test instead of a full method. Where a backend is named, the score belongs to GPT-Live and a second model together, and OpenAI names a backend only for GPT-Live-1, so some of the gap may come from the backend. And timing depends on where the clock is read. [InteractionBench](https://arxiv.org/abs/2611.02430), a paper posted to arXiv on November 2, 2026, measures timing where the client hears each system, so hosted systems include network delay. GPT-Live’s server reports its own clock, which let the authors remove about 1.5 seconds of delay; GPT-Live’s timing score then fell from 86% to 56%. Two latency figures are comparable only when both say where they were measured, and OpenAI’s chart gives the interval but not the point. In our reading, the latency and interactivity gaps are the results to take most seriously, because no backend sits behind them, though even they still need an outside run that states where the clock was read. The tool-calling and Tau3 gaps do not yet show that the voice model itself is better at tasks. They show what GPT-Live-1 does with GPT-6 Astra or Terra behind it, against Realtime models for which OpenAI names no backend. The launch post also quotes customers. Speak, the language-learning app, found in early evaluations that GPT-Live-1 gave learners more time to think before the tutor replied and cut interruptions by nearly 80% compared with a conventional turn-based system. Tony Stoyanov, co-founder and CTO of EliseAI, said GPT-Live-1 let the company simplify its codebase by 80% and remove 23,000 lines of code compared with its cascaded system. Both are the customers’ own figures, as OpenAI published them. ## 6. Pricing, phone calls and voices GPT-Live-1’s price covers the voice layer only. OpenAI’s [pricing page](https://developers.openai.com/api/docs/pricing) lists $0.05 per minute, billed per second, with backend model and tool usage charged separately. Capacity is counted in concurrent sessions: 50 on the Build tier, 300 on Launch and 500 on Grow, with the Free tier unsupported. The Realtime models are still billed by the token and limited by requests and tokens per minute; [gpt-realtime-2.1](https://developers.openai.com/api/docs/models/gpt-realtime-2.1) allows 400 requests and 200,000 tokens per minute on Build. Comparing the two schemes takes an assumption about how many tokens a minute of speech uses, and OpenAI published that conversion only for its 2024 preview model. For a phone deployment we would treat $0.05 a minute as the floor of what a call costs. The backend decides the rest, and it is the part of the bill that varies. *Figure: How OpenAI bills realtime voice.* | Model | Launched | Billed by | Price | | -------------------------------- | ---------------- | --------------- | ------------------------------------------------------------------------------ | | gpt-4o-realtime-preview | Oct 2024 | Audio tokens | At launch: $100 in, $200 out per 1M (OpenAI: about $0.06 and $0.24 per minute) | | gpt-realtime | Aug 2025 | Audio tokens | $32 in ($0.40 cached), $64 out per 1M | | gpt-realtime-2, gpt-realtime-2.1 | May and Jul 2026 | Audio tokens | $32 in ($0.40 cached), $64 out per 1M | | gpt-realtime-2.1-mini | Jul 2026 | Audio tokens | $10 in ($0.30 cached), $20 out per 1M | | gpt-realtime-translate | May 2026 | Minutes | $0.034 per minute | | gpt-realtime-whisper | May 2026 | Minutes | $0.017 per minute | | gpt-live-1 | Sep 2026 | Session seconds | $0.05 per minute, backend billed separately | Audio prices from OpenAI’s launch posts and pricing page; the preview row is its October 2024 launch price. Text and image tokens on the Realtime models, and GPT-Live-1’s backend model and tool usage, are charged separately. Calls run on SIP, which the Realtime API has accepted since August 2025. For inbound calls to GPT-Live, a developer points a SIP trunk at OpenAI’s SIP address and accepts each call through a webhook. Since November 2, 2026 any organization can also [place outbound calls](https://developers.openai.com/api/docs/guides/voice-sip). The application creates a Live session with SIP transport and supplies its own trunk, which must support TLS signaling and SDES-SRTP media, plus a caller number. Ringing is limited to three minutes and a connected call to two hours. Outbound calling is not offered through the Realtime API’s call-creation endpoint. The GPT-Live guide also covers partner routes through LiveKit, Twilio, Telnyx, Bandwidth and Daily with Pipecat. The voice list grew twice. At general availability OpenAI introduced twelve new voices, ten in English with Australian, British, Irish, North American, Filipino and Southern U.S. influences and two in Brazilian Portuguese. Nine more in German, French, Hindi, Japanese and Korean followed on October 16, 2026, which brings the [table of additional voices](https://developers.openai.com/api/docs/guides/live-conversations) to 21, alongside the default voice, marin. [Prompt-to-voice](https://developers.openai.com/api/docs/guides/custom-voices) opened to all API customers the same day. A developer describes a voice in text, through the `/v1/audio/voices` endpoint or the voice Playground, and gets back a saved voice for GPT-Live sessions. Creating one is free, and each organization can keep up to 100. The voices work only with GPT-Live, not with the text-to-speech, Realtime or Chat Completions APIs. OpenAI’s rules require adult voices, disclosure that the speech is AI-generated and preserved provenance metadata or watermarks, and they rule out asking for an imitation of a named person, public figure or fictional character. Voices made from recordings are a separate process for eligible customers, enabled through sales, and need a consent recording from the voice’s owner. Since July 31, 2026, supported GPT-Live audio from ChatGPT Voice and the API has carried SynthID watermarks, and OpenAI offers a public verification tool and an API for checking provenance. In ChatGPT, GPT-Live uses preset voices with safeguards against imitating real people. OpenAI also points enterprises that deploy AI agents to Presence, which it describes as another way to build voice workflows on GPT-Live-1; access goes through OpenAI’s account directors. ## 7. What OpenAI has not published As of November 10, 2026, OpenAI’s pages give no supported-language list for GPT-Live-1. For ChatGPT, OpenAI says it optimized GPT-Live for several of the most-used languages and warns that in some languages the model may have a non-native accent or gaps in fluency. The API voice table covers seven languages, and OpenAI says more voices and languages will follow over the coming months. The missing language list surprises us most. OpenAI added voices in German, French, Hindi, Japanese and Korean on October 16, 2026, yet a developer still cannot read which languages the model is meant to handle, and has to test each one. The API model takes audio and text only. Image and video input are not supported, and neither are structured outputs or fine-tuning; the knowledge cutoff is July 31, 2025. GPT-Live-1 mini, the Free-tier voice model in ChatGPT, has no API model page or changelog entry, so developers get one size. We found no reproduction of the Full-Duplex-Bench results outside OpenAI under stated conditions. The full cost of a call is $0.05 a minute plus whatever the backend consumes, and OpenAI publishes no figure for a typical backend share. ## 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 (August 3, 2026)](https://openai.com/index/continuous-voice-interaction-with-gpt-live/) 3. [OpenAI, Build more natural voice experiences with GPT-Live-1 in the API (September 10, 2026)](https://openai.com/index/introducing-gpt-live-1-in-the-api/) 4. [OpenAI API changelog](https://developers.openai.com/api/docs/changelog) 5. [OpenAI, GPT-Live 1 model page](https://developers.openai.com/api/docs/models/gpt-live-1) 6. [OpenAI, Getting started with GPT-Live](https://developers.openai.com/api/docs/guides/live) 7. [OpenAI, GPT-Live conversations and voice options](https://developers.openai.com/api/docs/guides/live-conversations) 8. [OpenAI, SIP guide](https://developers.openai.com/api/docs/guides/voice-sip) 9. [OpenAI, Custom voices](https://developers.openai.com/api/docs/guides/custom-voices) 10. [OpenAI, Decisions in voice](https://developers.openai.com/api/docs/guides/decisions-voice) 11. [OpenAI API pricing](https://developers.openai.com/api/docs/pricing) 12. [OpenAI, GPT-Realtime-2.1 model page](https://developers.openai.com/api/docs/models/gpt-realtime-2.1) 13. [OpenAI API deprecations](https://developers.openai.com/api/docs/deprecations) 14. [OpenAI, Advancing voice intelligence with new models in the API (May 7, 2026)](https://openai.com/index/advancing-voice-intelligence-with-new-models-in-the-api/) 15. [OpenAI, Introducing gpt-realtime and Realtime API updates (August 28, 2025)](https://openai.com/index/introducing-gpt-realtime/) 16. [OpenAI, Introducing the Realtime API (October 1, 2024, with later updates)](https://openai.com/index/introducing-the-realtime-api/) 17. [OpenAI, Hello GPT-4o (May 13, 2024)](https://openai.com/index/hello-gpt-4o/) 18. [OpenAI, ChatGPT can now see, hear, and speak (September 25, 2023)](https://openai.com/index/chatgpt-can-now-see-hear-and-speak/) 19. [Lin et al., Full-Duplex-Bench (arXiv 2503.04721)](https://arxiv.org/abs/2503.04721) 20. [He et al., InteractionBench (arXiv 2611.02430)](https://arxiv.org/abs/2611.02430) --- _Originally published at [https://www.fullduplex.ai/blog/v15-openai](https://www.fullduplex.ai/blog/v15-openai)._ _Part of **The Verticals** · v15 / 26 · from Fullduplex._ _Full index: https://www.fullduplex.ai/blog · Markdown of every article: https://www.fullduplex.ai/llms-full.txt._