What is GPT-Live-1 (and GPT-Realtime-2.1)?
GPT-Live-1 is OpenAI's voice model for live conversation, the default ChatGPT Voice model since July 2026 and available in the API since 10 September. It handles listening, speaking and turn-taking in one model, and it is full-duplex: it keeps listening while it talks, so callers can interrupt naturally. It does not do the reasoning itself. You attach a text model, such as GPT-6 Astra or GPT-6 Luna or a third-party model, to handle thinking and tool calls while the voice keeps the conversation going.
It ships with 12 voices, keyword biasing and better recognition of order numbers and postcodes, and supports phone calls. It takes audio and text only, with no image input, through the v1/live/sessions endpoint.
It costs $0.05 per session minute, billed per second, plus the attached text model's tokens. OpenAI's turn-based alternative, GPT-Realtime-2.1, is priced per token: $32 per million audio input tokens and $64 output, with a cheaper mini version at $10 and $20, and accepts images. GPT-Realtime-Translate, a live speech translator, costs $0.034 a minute.
How much does GPT-Live-1 (and GPT-Realtime-2.1) cost?
| Plan | Price | What you get |
|---|---|---|
| GPT-Live-1 | $0.05 per session minute | Plus backend model and tool usage |
| GPT-Realtime-2.1 | $32 in / $64 out per 1M audio tokens | Cached input $0.40; accepts images |
| GPT-Realtime-2.1 mini | $10 in / $20 out per 1M audio tokens | |
| GPT-Realtime-Translate | $0.034 a minute | Speech-to-speech translation |
GPT-Live-1 $0.05 per session minute, billed per second, plus the backend model's tokens. GPT-Realtime-2.1 audio $32 input / $64 output per 1M tokens ($0.40 cached); mini $10 / $20. GPT-Realtime-Translate $0.034 a minute.
Is GPT-Live-1 (and GPT-Realtime-2.1) worth it?
GPT-Live-1's improvement is conversational, not intellectual. Because it listens while speaking, it handles the pauses, interruptions and overlapping speech that make turn-based voice agents feel robotic. OpenAI reports a 30-point gain on Full Duplex Bench over GPT-Realtime-2.1, though it published only the gap, and Speak, a customer, reports 80% fewer interruptions. These are OpenAI's and customers' figures.
The design separates voice from reasoning. The voice layer costs a flat $0.05 a minute, $3 an hour, and you pick the text model behind it: a cheap one for order updates, GPT-6 Astra for complicated complaints. That is flexible, but it means the per-minute price is only part of the cost. OpenAI's top result on the Tau3 agent benchmark comes from pairing it with Astra at medium effort, its most expensive model.
GPT-Realtime-2.1 remains the choice when the agent needs to see images, which GPT-Live-1 dropped. It costs roughly $0.05 to $0.11 a minute in typical use, and its mini version about $0.016. GPT-Realtime-Translate, at $0.034 a minute, translates speech from over 70 input languages into 13 output languages as it arrives.
The gaps are latency figures, which OpenAI has not published, and image input. GPT-Live-1.1 is already rolling out in ChatGPT and listed as coming to the API.
Where it earns its place
A retailer running customer-service phone lines
Handle order and delivery calls without awkward interruptions
Pair GPT-Live-1 with GPT-6 Luna for routine calls. 1,000 hours of calls cost $3,000 for the voice layer, plus Luna's tokens; route escalations to Astra.
A team whose agent needs to see what the caller sends
Discuss a photo or screenshot during a call
Use GPT-Realtime-2.1, which accepts images, at about $0.05 to $0.11 a minute; or the mini version at roughly $0.016 for simpler calls.
A company running multilingual events or support
Translate live speech between languages
GPT-Realtime-Translate streams translated audio and transcripts from over 70 input languages into 13 output languages at $0.034 a minute.
Before you commit
- Setup
- Create a session on v1/live/sessions with gpt-live-1, choose a voice and attach a backend text model for reasoning and tools; GPT-Realtime-2.1 uses the Realtime API instead.
- Integrations
- OpenAI Realtime and Live APIs, function calling, telephony support; any OpenAI or third-party text model as the backend. GPT-Live-1 is the default ChatGPT Voice for Go, Plus and Pro.
- Data and privacy
- OpenAI does not train on API data unless you opt in; abuse-monitoring logs are kept up to 30 days. Data residency endpoints cost 10% more for models released after March 2026.
- Cost at scale
- Voice layer $3,000 per 1,000 hours of calls, before the backend model's tokens. GPT-Realtime-2.1 runs about $3 to $6.60 an hour in typical use; its mini version about $1.
- Limits
- Audio and text only; 12 voices, custom voices through sales. 25 concurrent sessions at Tier 1, 500 at Tier 5. Knowledge cutoff 31 July 2025. GPT-Realtime and GPT-Realtime-mini shut down on 20 January 2027.
What's good
- Full-duplex: listens while talking, so interruptions work naturally
- Flat $0.05 a minute for the voice layer
- Any text model can handle reasoning and tools behind it
- Up to 500 concurrent sessions at Tier 5
What's not
- No image input, unlike GPT-Realtime-2.1
- No published latency figures
- Total cost depends on the backend model you attach
- Not available on the free API tier
Best for
Teams building high-volume phone support or sales agents who want natural turn-taking and freedom to choose the reasoning model.
GPT-Live-1 (and GPT-Realtime-2.1) alternatives
GPT-Live-1 (and GPT-Realtime-2.1): common questions
- Is GPT-Live-1 (and GPT-Realtime-2.1) worth it?
- OpenAI's most natural-sounding way to build a phone agent, with interruptions handled properly; the bill is simpler per minute but not necessarily smaller. We rate it 4.3 out of 5.
- How much does GPT-Live-1 (and GPT-Realtime-2.1) cost?
- GPT-Live-1 $0.05 per session minute, billed per second, plus the backend model's tokens. GPT-Realtime-2.1 audio $32 input / $64 output per 1M tokens ($0.40 cached); mini $10 / $20. GPT-Realtime-Translate $0.034 a minute.
- What are the downsides of GPT-Live-1 (and GPT-Realtime-2.1)?
- No image input, unlike GPT-Realtime-2.1; no published latency figures; total cost depends on the backend model you attach; not available on the free API tier.
- Who is GPT-Live-1 (and GPT-Realtime-2.1) best for?
- Teams building high-volume phone support or sales agents who want natural turn-taking and freedom to choose the reasoning model.
- How much does GPT-Live-1 (and GPT-Realtime-2.1) cost for a team or in production?
- Voice layer $3,000 per 1,000 hours of calls, before the backend model's tokens. GPT-Realtime-2.1 runs about $3 to $6.60 an hour in typical use; its mini version about $1.
- Is GPT-Live-1 (and GPT-Realtime-2.1) safe for company data?
- OpenAI does not train on API data unless you opt in; abuse-monitoring logs are kept up to 30 days. Data residency endpoints cost 10% more for models released after March 2026.
- What are the best GPT-Live-1 (and GPT-Realtime-2.1) alternatives?
- The closest alternatives we have reviewed are Eleven v4 and v4 Turbo, Voxtral 3 (Transcribe and TTS), Gemini 3.8 Flash TTS and Cartesia Sonic 3.6.
Answers checked Oct 25, 2026.
Make GPT-Live-1 (and GPT-Realtime-2.1)? Show this rating on your site
