Live voice
GPT Live 1
Real-time voice conversations with low latency and support for interruptions.
gpt-live-1Best for: Real-time voice agents and phone-style calls
Catalog status
Enabled
Research checked
2026-10-06
Benchmark results
Results are source-attributed, not TokenRex measurements. Different tests and thinking variants are not interchangeable.
2 of 2 results
Tau3 voice tasks (airline/retail/telecom), paired with GPT-6 Astra
Accuracy · Third-party report
Test conditions and source
Vendor-reported launch benchmark; GPT-Live-1 delegates reasoning to GPT-6 Astra (medium effort) backend — this is a pipeline score, not GPT-Live-1 in isolation. Compared only against OpenAI's own prior model, not an independent/neutral arena. Published source text: OpenAI's official launch post (mirrored on community forum) states: 'Paired with GPT-6 Astra at medium reasoning effort, GPT-Live-1 completed 83.6% of Tau3 tasks on the first attempt, versus 45.7% for GPT-Realtime-2.1.'
Tested model: gpt-live-1
OpenAI Developer Community — Introducing GPT-Live-1 in the API · Checked 2026-10-06
Full Duplex Bench v1 — turn-taking latency
Accuracy · Third-party report
Test conditions and source
Lower is better; vendor-reported, compared only to OpenAI's own previous Realtime model. Published source text: Official launch post: '0.798-second turn-taking latency, compared with 1.41 seconds for GPT-Realtime-2.1.'
Tested model: gpt-live-1
OpenAI Developer Community — Introducing GPT-Live-1 in the API · Checked 2026-10-06
Model specifications
Input and output modalities
Provider-reported formats; API compatibility is detailed below.
Input
Output
- Exact model ID
- gpt-live-1
- Model type
- Live voice
Exact token limits are not published in the reviewed sources, so no estimated limits are shown.
Versions and thinking levels
One enabled version is currently listed for this model.
GPT Live 1 — everything it can do
GPT Live 1 is OpenAI's full-duplex voice model: it listens and speaks at the same time, handles interruptions naturally, and hands reasoning and tool calls to a backend agent. On TokenRex you pay the official price — $0.05 per minute, billed per second with no round-up.
How it works
Example: the user asks "Where is my order?" GPT Live keeps the conversation going while your backend checks the order database, then answers naturally when the result arrives. The backend can be a supported OpenAI Responses model or your own external agent.
Capabilities
| Speech input (microphone) | Yes |
| Speech output (natural voice) | Yes |
| Real-time two-way conversation | Yes |
| Full duplex — listens while speaking | Yes |
| User interruption / barge-in | Yes |
| Silence handling | Yes |
| Background-noise handling | Yes |
| Native speech transcript | Yes |
| Response text alongside audio | Yes |
| Multilingual conversation | Yes |
| Translation | Yes(via your backend setup) |
| Function calling (your tools) | Yes |
| Backend model delegation | Yes |
| Web / tool use | Yes(through delegated backend) |
| Phone calls / telephony (SIP) | Yes |
| WebRTC | Yes |
| WebSocket | Yes |
| Long conversations | Yes |
| Image input directly | No |
| Video input directly | No |
| Image generation directly | No |
| Video generation directly | No |
| Fine-tuning | No |
| Structured outputs | No |
Voices
All 22 voices below opened a real GPT Live 1 session in our tests, including the classic OpenAI voices (alloy, ash, coral, echo, sage, shimmer, verse, ballad, cedar). Set it with session.audio.output.voice; the default is marin and the voice is fixed for the call. Voice and language are separate: any voice can speak Hindi, Hinglish or another language — put the rule in session.instructions (for example, "Speak casual Hinglish unless the caller asks to switch"). The accent column is only the voice's natural accent. Signed-in users can try every voice and language in Playground → Talk live.
| Voice | Language / accent | Character |
|---|---|---|
| Marin (marin) | Neutral (default) | Feminine |
| Cedar (cedar) | Neutral | Masculine |
| Alloy (alloy) | Neutral | Neutral |
| Ash (ash) | Neutral | Masculine |
| Ballad (ballad) | Neutral | Masculine |
| Coral (coral) | Neutral | Feminine |
| Echo (echo) | Neutral | Masculine |
| Sage (sage) | Neutral | Feminine |
| Shimmer (shimmer) | Neutral | Feminine |
| Verse (verse) | Neutral | Masculine |
| Quartz (quartz) | English / Australian | Feminine |
| Ripple (ripple) | English / Australian | Masculine |
| Vesper (vesper) | English / British | Masculine |
| Willow (willow) | English / Irish | Feminine |
| Stone (stone) | English / Irish | Masculine |
| Gleam (gleam) | English / North American | Feminine |
| Meridian (meridian) | English / North American | Masculine |
| Bossa (bossa) | Portuguese / Brazilian | Feminine |
| Tempo (tempo) | Portuguese / Brazilian | Masculine |
| Beacon (beacon) | English / Filipino | Masculine |
| Delta (delta) | English / Southern US | Feminine |
| Cinder (cinder) | English / Southern US | Masculine |
Pricing — at market price, billed per second
Official price is $0.05 per minute. TokenRex charges the same: $0.05 per minute, billed per second — a 20-second call costs about $0.017, not a full minute. This covers the Live voice layer only; the backend reasoning model and any tools it calls are billed at their own normal market usage prices.
| Usage | Market price | TokenRex (at market price) |
|---|---|---|
| 1 minute | $0.05 | $0.05 |
| 10 minutes | $0.50 | $0.50 |
| 30 minutes | $1.50 | $1.50 |
| 1 hour | $3.00 | $3.00 |
| 10 hours | $30 | $30 |
| 100 hours | $300 | $300 |
Tools and phone calls
- Function calling: register your own tools — calendar, orders, payments, database, CRM, search, booking, customer support or your own APIs and agents. GPT Live calls them and speaks the result.
- Phone agents: connect through SIP/telephony for customer support, reservations and other phone-call applications. Architecture: Phone call → SIP/Telephony → GPT Live 1 → backend tools → response → caller. Official partner integrations include LiveKit, Twilio, Telnyx and Daily/Pipecat.
- Full duplex: unlike a classic speech-to-text → LLM → text-to-speech pipeline, user and model can talk at the same time. If the user interrupts — "Wait, actually tell me the price" — the model handles it naturally instead of waiting for a rigid turn boundary.
Model ID:
gpt-live-1wss://tokenrex.ai/api/public/v1/live or WebRTC — see the Docs live-voice section for full connection code.Features and tool support
TokenRex’s exposed interface, not every feature advertised by the provider. “Available” describes the implemented interface, not a successful test of every input or tool.
Input through TokenRex
AvailableLive microphone audio over WebRTC; session start and control events over wss://tokenrex.ai/api/public/v1/live. Tested: sessions opened for all 22 voices.
Output through TokenRex
AvailableSpoken audio streamed back in real time, plus live transcript events for both sides of the call.
Function tools
AvailableSend your own tools with session.start; TokenRex forwards each call to your app and returns your verified result to the conversation.
Languages
AvailableAny voice can speak Hindi, Hinglish, English and other languages — set the rule in session.instructions.
Sources and verification
Provider specifications describe the upstream model, not a guarantee of every feature through TokenRex. Prices come from the enabled catalog; benchmark results belong to the exact tested model.
- Vendor model documentation
Vendor-reported
Last checked - OpenAI Developer Community — Introducing GPT-Live-1 in the API
Third-party report
Last checked
Research notes (6)
- Full-duplex voice model: listens and speaks simultaneously, with interruption (barge-in), silence and background-noise handling and a native speech transcript.
- Audio-only: no image or video input, no image or video generation, no fine-tuning, no structured outputs.
- Supports function calling and backend model delegation; web/tool use runs through the delegated backend. Reasoning in vendor benchmarks delegates to GPT-6 Astra.
- Connects over WebRTC, WebSocket or SIP; telephony partner integrations include LiveKit, Twilio, Telnyx and Daily/Pipecat.
- Voices: Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, Cinder; default marin; custom voices supported.
- Official price $0.05/minute billed per second; TokenRex charges 50% (₹2.40/minute). Backend model and tool usage is billed separately.