WebRTC
Default choice for browser or mobile clients that need low-latency audio and direct media handling.
Consumer product vs developer API
GPT-Live vs GPT-Realtime 2.1
GPT-Live is the consumer Voice experience inside ChatGPT. GPT-Realtime-2.1 and GPT-Realtime-2.1 mini are documented API models for building your own voice agents. Similar realtime interaction does not make them interchangeable products or model IDs.
Last verified:
The boundary that matters
| Question | GPT-Live | GPT-Realtime 2.1 |
|---|---|---|
| Where does it run? | Inside ChatGPT Voice, including desktop Work and Codex | Inside your app, service, or telephony workflow |
| Who controls the UI and account? | OpenAI / ChatGPT | You, using an OpenAI API project |
| Model names | GPT-Live-1 and GPT-Live-1 mini | gpt-realtime-2.1 and gpt-realtime-2.1-mini |
| Access model | ChatGPT plan, region, workspace, and rollout | API account, project limits, and documented endpoints |
| Pricing unit | ChatGPT subscription and Voice allowances | Token-based API pricing |
| Transport | Managed by ChatGPT | WebRTC, WebSocket, or SIP |
GPT-Realtime-2.1 models and prices
Per one million tokens. Both models list a 128,000-token context window and 32,000 maximum output tokens.
| Model | Text in / cached / out | Audio in / cached / out | Image in / cached |
|---|---|---|---|
| gpt-realtime-2.1 | $4 / $0.40 / $24 | $32 / $0.40 / $64 | $5 / $0.50 |
| gpt-realtime-2.1-mini | $0.60 / $0.06 / $2.40 | $10 / $0.30 / $20 | $0.80 / $0.08 |
WebRTC, WebSocket, or SIP?
WebSocket
Use for server-side applications, custom audio pipelines, and environments where your backend owns the session.
SIP
Use for inbound phone calls, contact centers, and existing telephony providers. Outbound dialing requires a provider-controlled dialer or media bridge.
Choose the right path
You only want to use Voice
Use GPT-Live in ChatGPT and choose a Voice mode that matches the task.
You are building a browser or mobile agent
Prototype with WebRTC and gpt-realtime-2.1; benchmark mini against your latency, quality, and cost targets.
You are building a phone agent
Use SIP or a server-side media bridge, add webhook verification, consent, escalation, and transfer controls.
You specifically need GPT-Live-1
Join the notification list and wait for official API access; do not invent a model ID or treat the form as an API launch.
GPT-Live vs Realtime FAQ
Can I send gpt-live-1 as an API model name?
No documented public API contract supports that model ID as of 2026-07-19. Use the official Realtime model IDs.
Do GPT-Realtime-2.1 models support images?
Yes for image input. Their model pages list text, audio, and image input; text and audio output; and no video.
Is mini just for prototypes?
No. Mini is a lower-cost, faster model that can be production-appropriate when testing shows its quality meets the task.
Official sources
OpenAI Help Center · verified ChatGPT VoiceOpenAI · verified GPT-Live-1 in the APIOpenAI Developers · verified GPT-Realtime-2.1 modelOpenAI Developers · verified GPT-Realtime-2.1 mini modelOpenAI Developers · verified Realtime API with WebRTCOpenAI Developers · verified Realtime API with WebSocketOpenAI Developers · verified Realtime API with SIPOpenAI · verified ChatGPT Voice in Work and Codex