AIREITERA launch post can make a model look available before developers can actually call it. For GPT-Live-1, OpenAI’s public API signup page still asks builders to get notified when the model becomes available, while the documented Realtime API has published rates. That distinction matters: do not put GPT-Live-1 into a production budget or integration until OpenAI publishes a model ID, endpoint, documentation, and rate card.
The decision in one minute
GPT-Live-1 is not publicly documented as a generally available API in the OpenAI material I could verify. OpenAI’s GPT-Live-1 API page is a notification form, not an API reference: it provides no endpoint, model ID, pricing, limits, or launch date.
That does not prove that no private access exists. It does mean a public developer cannot treat a social announcement or a ChatGPT voice experience as a callable API contract.
For a production voice application today:
| Need | Practical choice |
|---|---|
| Ship a documented realtime voice API now | GPT-Realtime-2.1 |
| Reduce audio and text-token cost | GPT-Realtime-2.1 mini |
| Prototype future model switching | Put the model ID and pricing in configuration |
| Require GPT-Live-1 specifically | Join the notification list; do not promise a launch date |
GPT-Live-1 pricing is not a published API price
The verified OpenAI signup page says, “Sign up to get notified when GPT‑Live‑1 is available in the API.” It does not list a per-minute price, audio-token price, text-token price, minimum charge, or usage limit. As a result, the honest GPT-Live-1 API price is unknown, not $0.05 per minute and not any estimate copied from another realtime model.
ChatGPT subscription pricing is a separate product surface. A Plus or Pro subscription may include consumer voice access, but it does not automatically provide API credits. Developers should keep the subscription invoice and API invoice in separate budgets.
What you can buy and build today
OpenAI’s documented GPT-Realtime model page publishes token-based rates. The figures below are the rates found in the research materials for the current Realtime family; check the official page again before deployment because rates can change.
| Model | Text input / cached | Text output | Audio input / cached | Audio output | Context / max output |
|---|---|---|---|---|---|
| GPT-Realtime-2.1 | $4 / $0.40 per 1M | $24 per 1M | $32 / $0.40 per 1M | $64 per 1M | 128k / 32k |
| GPT-Realtime-2.1 mini | $0.60 / $0.06 per 1M | $2.40 per 1M | $10 / $0.30 per 1M | $20 per 1M | 128k / 32k |
Both models are described in the researched API-status material as accepting text, audio, and image input and producing text and audio output; video is not listed as supported. The standard model is the safer choice when complex reasoning or tool use matters. Mini is the cost-first choice when latency and price matter more than maximum capability.
The important billing detail is that input and output are separate. A five-minute call is not automatically five minutes of user audio plus five minutes of assistant audio. Measure how much each side actually speaks.
Build a budget that can survive a model change
Until GPT-Live-1 publishes a rate card, use a scenario model rather than a guessed blended price:
monthly exposure = sessions
× measured billable usage per session
× provider rate
× retry and tool-call factorsTrack these fields in every pilot:
- User speech seconds and assistant speech seconds separately.
- Text input, cached input, and text output tokens.
- Turns, tool calls, failed tools, retries, and reconnects.
- Session duration, silence, abandoned sessions, and peak concurrency.
- Fallback traffic, such as text-only mode or a smaller model.
For example, a planning worksheet can use 500 input minutes and 500 output minutes, but it must label any audio-token-per-minute conversion as an assumption. Exact budgeting should use the usage records returned by the API, not a calculator’s generic minute conversion.
Set a hard session cap, a tool-call cap, a spend alert, and a kill switch before opening a pilot. A voice agent that reconnects after a network drop can create usage without completing the customer’s task, so “fair use” language is not a financial control.
GPT-Live-1 API FAQ
Does the notification form grant GPT-Live-1 API access?
No. The public form collects contact and business details so OpenAI can notify interested builders. It does not display an API key, model ID, endpoint, pricing contract, or guaranteed beta enrollment.
Can ChatGPT Plus or Pro pay for API usage?
No. ChatGPT subscriptions and API usage are billed separately. Choose a consumer plan for the ChatGPT product; budget developer API calls from the API rate card.
Is GPT-Live-Transcribe the same as GPT-Live-1?
No. A transcription model converts speech to text. GPT-Live-1 refers to the conversational voice experience. A transcription-minute price cannot be used as the price of a full-duplex agent that listens, reasons, speaks, and potentially calls tools.
Should a team wait or use Realtime now?
Use GPT-Realtime-2.1 or its mini variant for production now. If GPT-Live-1’s interaction model is strategically important, isolate the model configuration behind an adapter so a later migration changes configuration rather than your entire application.
“Much more effective to dictate the same prompt within a chatgpt work window.” — @pranavkrn, describing a limitation he observed with voice tool calling; original post. This is a user report about the ChatGPT experience, not evidence about an undocumented API.
The practical call
The cleanest decision is to ship on a documented Realtime model, record the usage categories that determine cost, and keep GPT-Live-1 as a replaceable configuration target. Waiting is justified only when GPT-Live-1’s specific full-duplex behavior is a hard product requirement. Otherwise, an unpublished price and endpoint create avoidable delivery and margin risk.