OpenAI made GPT-Live-1 generally available in its API on September 10, 2026. It is a voice model that listens and speaks at the same time, and a voice session costs $0.05 per minute, billed per second. Reasoning, tools and the phone line are paid apart. For AI phone agents, the GPT-Live API changes how natural a call can sound. It does not run a calling operation. We read the OpenAI docs on November 19, 2026, and every OpenAI claim below links to them.
Key Takeaways
- OpenAI released GPT-Live-1 to general availability in its API on September 10, 2026, through the v1/live/sessions endpoint.
- A GPT-Live voice session costs $0.05 per minute, billed per second, and OpenAI bills backend model and tool usage separately.
- OpenAI's model page lists speed as "Not specified", so GPT-Live has no published latency figure to plan a phone campaign around.
- Since November 2, 2026, outbound SIP calling through the Live API is open to all OpenAI organizations, using your own SIP provider.
- Topcalls AI calls start from $0.25 a minute, depending on the country called, billed by the second, with recording, transcription and analytics included.
What is the GPT-Live API?
GPT-Live is OpenAI's voice model for full-duplex conversation, meaning it can listen while it speaks. OpenAI's model page calls GPT-Live-1 "our premier model for natural, expressive voice conversations with smooth interruption handling". It runs on one endpoint, v1/live/sessions, and takes audio and text in and out.
OpenAI built it in two parts. In its GPT-Live guide, OpenAI writes: "GPT-Live handles conversation. It listens, speaks, and decides when to ask the backend for help." A separate backend model or agent does the reasoning and the tool work.
You pick the backend in one of two ways. With Responses delegation, OpenAI runs a model you choose. With client delegation, your own application runs the model, agent or workflow and returns the result. You set the mode when you create the session, and changing it means starting a new session.

You connect over WebRTC, WebSocket or SIP. A WebSocket sideband lets your server watch or steer a session that is already running. OpenAI's partner guide lists LiveKit, Twilio, Telnyx, Bandwidth and Daily/Pipecat as ways to bring existing calls or audio streams into GPT-Live.
How much does the GPT-Live API cost?
A GPT-Live voice session costs $0.05 per minute, billed per second, according to OpenAI's September 10 changelog entry. The same entry says backend model and tool usage is charged separately. That makes $0.05 the price of the voice layer, not of a finished phone call.
| Point | GPT-Live-1 (OpenAI docs) | Topcalls (product page) |
|---|---|---|
| Published price | $0.05 per minute, billed per second | From $0.25 a minute, depending on the country called, billed by the second |
| What it covers | The voice session | The AI voice model, call recording, transcription, analytics and the phone line |
| Billed separately | Backend model and tool usage | No seat fees for AI calling, no minute bundles, no separate transcription bill |
| Phone connection | Your SIP provider or a telephony partner | Included in the call price |
| Result of each call | Not listed on the model page | Recording, transcript, summary, disposition and post-call analysis |
A worked example shows why the two numbers are not rivals. Ten thousand minutes of GPT-Live voice sessions is $500 at $0.05. Ten thousand minutes of Topcalls AI calls at the $0.25 starting price is $2,500. The first figure leaves out the backend model, the phone line and the code you write around it. We cannot price those, because they depend on your choices.
Want the numbers for your own call volume? Run the ROI calculator, or read our voice agent cost breakdown.
Can GPT-Live make phone calls?
Yes. GPT-Live can take and place phone calls through SIP. OpenAI's changelog says that since November 2, 2026, outbound SIP calling is available to all organizations through the Live API, and that you connect GPT-Live to a phone number through your SIP provider.
A session is one conversation. A campaign is thousands of them. The model page and the guides describe the first and say nothing about the second. A campaign still needs these parts:
- A lead list and a schedule: who gets called, and in which hours.
- Retries: what happens on a busy line or no answer, and when the next attempt goes out.
- Concurrency: how many calls run at once, and what happens at the limit.
- A record of each call: a recording, a transcript, a summary and an outcome you can report on.
- Results in your systems: a way to push outcomes to your CRM through an API and webhooks.
Topcalls covers these in smart campaigns. Each account includes 20 concurrent outbound calls, and extra slots can be added. The platform controls dispatch hours and retries per campaign. The customer owns consent and opt-outs, and a better voice does not change that.
Does a realtime voice model fix AI phone call latency?
It helps with turn-taking, but OpenAI publishes no latency figure to prove more. The model page lists performance and speed as "Not specified", and the guide describes full-duplex behavior without numbers. Treat any latency claim about GPT-Live as unverified until you measure it on your own phone line.
On a phone call, the delay a caller hears is more than the model. It includes the telephony hop and, with GPT-Live, the wait when the voice model hands a question to a backend model. That wait depends on the backend you choose and the tools it calls.

Our guide on AI voice latency for outbound calls lists what to test before you buy. The same test applies to a GPT-Live build. Our barge-in glossary entry explains the interruption behavior that full-duplex models aim to improve.
What does GPT-Live change for AI calling?
GPT-Live raises the floor for how a voice agent sounds, and it gives builders a price to plan around. Here is what changes for phone work:
- Interruptions: OpenAI built the model around smooth interruption handling, which is the moment most scripted voice agents break.
- Custom voices: on October 16, 2026, OpenAI released prompt-to-voice, which creates a custom voice from a text description. Creating one is free, and speech uses regular GPT-Live session billing.
- Phone access: SIP and the telephony partners mean a developer can attach a number without a separate speech stack.
None of these decides whether a call books a meeting. That comes from the script, the data you feed the agent, the offer and the follow-up. Our post on how AI voice agents work covers those layers.
Should you build on GPT-Live or use an AI calling platform?
Build on GPT-Live if the voice experience is your product and you have engineers to own it. Use a calling platform if your goal is booked meetings, qualified leads or reactivated customers and you want a team to run the campaign layer. The build-or-buy question is the same one we cover in build vs buy for an AI voice agent.
Where a GPT-Live build does not fit: a sales team with no developers, a launch date inside a few weeks, or a need for call reports from day one. In those cases the work around the model takes longer than the model.
Topcalls is the other path. A dedicated team configures the agent and the campaign, and you measure it on your own KPIs. Our AI voice agents handle appointment setting and lead qualification in up to 78 languages. How many you get depends on the transcription and voice models you choose.
To see a campaign priced for your volume, book a strategy call.
Sources
- OpenAI API changelog: the September 10, October 16 and November 2, 2026 entries.
- GPT-Live-1 model page: description, endpoint, price and the "Not specified" speed line.
- GPT-Live guide: the two-part design, connections and billing.
- GPT-Live delegation guide and partner integrations.
Frequently Asked Questions
Get AI calling tips in your inbox
No spam. One email per week with actionable sales automation tips.






