OpenAI released GPT-Live-1 in the API on September 10, 2026. The launch post’s own summary: it “brings ChatGPT’s natural, full-duplex conversations to the API, with more control over how voice agents speak and act.” The model listens and speaks at the same time, and hands anything that needs thinking to a backend you pick.
This page is the record: what shipped, what it costs, how the delegation works, the benchmarks OpenAI published, and what it does and does not replace. Every quotation is from OpenAI’s launch post, the model page or the GPT-Live guides, all read the day of release.
On this page · 8 sectionsOpen
OpenAI released GPT-Live-1 in the API on September 10, 2026. It is a full-duplex voice model, ‘capable of listening and speaking at the same time’, first shipped inside ChatGPT and now sold to developers at $0.05 per minute, billed per second.
The architecture is the story. GPT-Live-1 only runs the conversation; it ‘can delegate reasoning and tool calls to a backend text model like GPT-6 Astra or a third-party model’, and OpenAI’s docs say plainly that ‘Developers choose the models, tools, and agent harness behind the conversation.’
Two delegation modes: Responses delegation, where GPT-Live calls an OpenAI model you configure, and client delegation, where your application runs any model, agent or service and returns results. In both, ‘Your application owns permissions, confirmations, private function execution, and durable task state.’
It replaces the chain. ‘Traditional voice agents stitch together speech-to-text, a reasoning model, and text-to-speech.’ One customer’s CTO says the single model ‘simplified our code base by 80% and removed 23K lines of code’; Speak reports interruptions cut by almost 80 percent versus turn-based systems.
Benchmarks, all OpenAI’s own: 30 percentage points better than GPT-Realtime-2.1 on Full Duplex Bench, and, paired with GPT-6 Astra at medium reasoning, first on Tau3, ‘which measures frontier voice-agent intelligence on end-to-end tasks’.
Pricing changes shape, not just level: the voice layer is a per-second clock while the Realtime models are still sold per audio token, $32 per million in and $64 per million out for gpt-realtime-2.1. Backend model and tool usage is billed separately, at the backend model’s normal rates.
It is not the Realtime replacement. OpenAI’s deprecations page still routes gpt-realtime to gpt-realtime-2.1 ahead of the January 20, 2027 shutdown, and GPT-Live-1 does not support the v1/realtime endpoint at all; it lives on v1/live/sessions. Twelve new voices ship with it.
§ 01What shipped
| Item | Detail |
|---|---|
| Model id | gpt-live-1 |
| Endpoint | v1/live/sessions (Chat Completions, Responses and Realtime endpoints all listed as not supported) |
| Modalities | Audio and text in, audio and text out; image and video not supported |
| Features | Streaming, function calling; no structured outputs, no fine-tuning |
| Knowledge cutoff | July 31, 2025 |
| Price | $0.05 per minute of session, billed per second; backend model and tools billed separately |
| Voices | Twelve new: Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, Cinder |
| Telephony | Full-duplex agents for phone calls, “from restaurant reservations to customer support” |
§ 02One model for the conversation, another for the thinking
OpenAI’s pitch begins with the architecture it is replacing. “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.” GPT-Live-1 collapses the voice half into one model and keeps the thinking half separate: it “can delegate reasoning and tool calls to a backend text model like GPT-6 Astra or a third-party model”, and “This lets the conversation continue while work happens in the background.”
The getting-started guide puts the two parts in one sentence each. GPT-Live “listens, speaks, and decides when to ask the backend for help.” The backend “reasons, uses tools, and returns results for GPT-Live to communicate. Keep detailed instructions, business rules, and tool workflows here.” And on the line that matters for anyone deploying this in a business: “Your application owns permissions, confirmations, private function execution, and durable task state.”
| Responses delegation | Client delegation | |
|---|---|---|
| Who runs the backend | GPT-Live calls the OpenAI Responses model you configure | Your application runs any model, agent harness or service |
| Who prepares context | GPT-Live supplies conversation context | Your application chooses which history, memory and state each request receives |
| When OpenAI recommends it | A configured model and tool loop fits the task | You need to validate, redact or combine results, use another backend, or route between code and models |
| What stays with you either way | Permissions, confirmations, business records, task state | The same |
| Switching | Chosen at session creation; a new session to change | The same |
§ 03The numbers OpenAI published
All of these are OpenAI’s evaluations on OpenAI’s benchmarks; no independent run existed on launch day.
| Claim | Source line |
|---|---|
| Full Duplex Bench | “improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1, with large gains in turn-taking latency and interactive behavior” |
| Tau3 | Paired with GPT-6 Astra at medium reasoning effort, “ranks #1 on Tau3, which measures frontier voice-agent intelligence on end-to-end tasks” |
| Speak | Learners got more time to think before the tutor answered, “cutting interruptions by almost 80% versus previous turn-based systems” |
| A healthcare customer | GPT-Live-1 “simplified our code base by 80% and removed 23K lines of code”, per its CTO |
| Yelp | Yelp Host and Hatch saw “meaningful improvements in call handling rates” and callers “speaking fuller, more natural sentences” |
§ 04Pricing: a clock for the voice, tokens for the thinking
The voice layer is $0.05 per minute, and the model page is specific: “Session duration is not rounded up to the next whole minute.” Ten minutes of conversation is $0.50 before the backend does anything. The backend is billed at its own rates: “Backend Responses calls use the normal pricing for the configured model and tools.” So the bill has two lines, and the second one is whatever agent you chose.
For contrast, OpenAI’s Realtime models, which put speech, reasoning and tool selection in one model, are still priced per audio token.
| Model | Audio input, per 1M tokens | Cached input | Audio output, per 1M tokens |
|---|---|---|---|
| gpt-realtime-2.1 | $32.00 | $0.40 | $64.00 |
| gpt-realtime-2.1-mini | $10.00 | $0.30 | $20.00 |
The two schemes do not convert cleanly, since a minute of speech is not a fixed number of audio tokens, and we will not pretend otherwise. What is true is that the GPT-Live-1 price is knowable before the call starts, and the backend cost scales with how much thinking the call needs.
§ 05What it does and does not replace
GPT-Live-1 is not listed as a Realtime replacement anywhere in OpenAI’s documentation. The deprecations page, updated July 20, 2026, retires gpt-realtime and gpt-4o-realtime on January 20, 2027 in favor of gpt-realtime-2.1. The getting-started guide draws the design line instead: “You choose the backend model or agent independently of the voice model; Realtime uses one model for speech, reasoning, and tool selection.” A separate transcription lineage is also moving: whisper-1, gpt-4o-transcribe and their siblings shut down February 26, 2027 in favor of gpt-live-transcribe or gpt-transcribe.
| Date | Event |
|---|---|
| July 20, 2026 | Deprecation notice for legacy audio, realtime and transcription models |
| September 10, 2026 | GPT-Live-1 generally available in the API |
| January 20, 2027 | gpt-realtime and gpt-4o-realtime shut down; gpt-realtime-2.1 is the replacement |
| February 26, 2027 | whisper-1 and the gpt-4o-transcribe family shut down; gpt-live-transcribe or gpt-transcribe replace them |
§ 06Where the agent lives
Our conflict, declared: we build CellCog, an AI employee platform, and it has no real-time voice channel today. Nothing on this page is a CellCog voice claim. We route Agent, Agent Creative and Agent Team to Claude Fable 5.1 at the Core and Max tiers and Gemini 3.8 Flash at Flash, none of which is a voice model.
What interests us is the split OpenAI just made official. A voice layer priced by the second, with “Developers choose the models, tools, and agent harness behind the conversation” written into the launch post, is OpenAI saying that the conversation and the worker are different things. The worker is where the permissions live, where the task state lives, where the memory of the last call lives, and where the business rules live; OpenAI’s own guide tells you to keep all of that in the backend. That backend is what an AI employee is: the part that remembers, holds the credentials, does the work, and asks before doing anything consequential. The voice on the phone is one more interface to it, next to chat, email and a task board.
§ 07What we are watching for
- Independent benchmarks. Every number above is OpenAI’s own; the first outside Full Duplex Bench or Tau3 run gets folded in here.
- The Agents API. OpenAI published “Introducing the Agents API” the same day; how it pairs with client delegation is its own record.
- Realtime’s future. Any deprecation-page change that names GPT-Live as a Realtime successor flips the “does not replace” line above.
- Voice and language expansion. OpenAI says more voices and languages come “over the coming months.”
§ 08Sources
OpenAI, Build more natural voice experiences with GPT-Live-1 in the API, September 10, 2026 (every quotation above is from this post, the GPT-Live 1 model page, the getting started and delegation guides, or the pricing and deprecations pages, all read September 10, 2026). CellCog routing statements reflect our configuration on the same date.
Q1What can the backend be?
With Responses delegation, a supported OpenAI Responses model such as GPT-6 Astra or Luna; OpenAI’s example pairs Luna with high-volume tasks like scheduling and Astra with complex customer issues. With client delegation, anything your application runs: your own model, a third-party model, an agent harness or an external service. The launch post includes a code sample connecting GPT-Live-1 to Codex.
Q2What does the application still have to own?
Permissions, confirmations, private function execution and durable task state, in either mode. The docs also warn that ‘Interrupting speech does not automatically cancel backend work’, so the backend needs its own lifecycle.
Q3How does it connect?
WebRTC for browser apps, WebSockets for server-side audio, sideband connections for backend control, and telephony through partner paths; the launch post names phone calls from restaurant reservations to customer support as a target. Input and output are audio and text; image and video input are not supported, and the knowledge cutoff is July 31, 2025.
Q4Which voices are new?
Twelve: Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta and Cinder, across accents, dialects and languages. Custom voices require contacting sales.
Q5Does CellCog have a voice interface?
Not today. CellCog employees work through chat, email, channels, tasks and the tools you connect; there is no real-time voice channel, and we make no claim about one here. What we do build is the half of this architecture OpenAI now describes as separate: the agent behind the conversation, with its own memory, permissions and task state.