Every comparison of realtime voice models measures the same things. Latency, benchmark scores, price per minute. All useful. None of them answer the question that actually blocks a phone deployment, which is whether the thing can pick up a call at all.
We support GPT-Live-1, Gemini Live and Grok Voice, so we went looking. The answer turned out to be less about capability and more about what each vendor is willing to put in writing.
Sockets are not phone lines
Here is the split as the vendors document it today.
OpenAI publishes a SIP connector for GPT-Live-1. Calls arrive at an endpoint shaped like sip:PROJECT_ID@sip.api.openai.com over TLS, and incoming calls route through a project webhook. Their GPT-Live guide lists telephony and SIP next to WebRTC and WebSockets as supported connection paths, and names LiveKit, Twilio, Telnyx and Daily as integration routes.
Google's Live API documentation describes a stateful WebSocket connection and stops there. xAI documents a WebSocket API plus a LiveKit integration. Neither documents SIP.
That does not mean Gemini and Grok can't reach a phone. It means the phone leg is yours to terminate rather than theirs. With GPT-Live-1 you point SIP at OpenAI and the carrier call lands there. With the other two you terminate the call yourself and feed audio into a socket, which is what our telephony layer already does. The risk is not the extra layer, it is that when the socket behaves oddly on a carrier call you are debugging a path the vendor never documented.
Documentation only takes you so far
We ran it. GPT-Live-1 over real telephony rather than browser sessions, carried by a carrier, every call completed. This is early, on a model only days old, so treat it as a first signal and not a reliability figure. What it settles is the thing docs alone cannot, which is that the path works end to end on a real call.
We get into the trace data and the edge cases that didn't fit here in the full write-up. Read it on Dograh.
The pricing trap that's easy to skim past
This one caught us out, and it will catch you out if you skim a pricing page.
GPT-Live-1 is $0.05 per minute, billed per second. That covers the voice layer. Reasoning and tool use delegate to a backend agent that bills separately, so your real number is $0.05 plus whatever that agent runs to. We saw this in our own run config: realtime_model: gpt-live-1 sitting next to llm_model: gpt-4.1, two line items for one call.
grok-voice-think-fast-2.0 is $0.08 per minute, which xAI also publishes as $4.80 per hour, with reasoning included.
gemini-3.8-live splits the meter. Audio in at $0.005 per minute, audio out at $0.018 per minute.
Look at those three again and notice they are not the same kind of number. One covers a voice layer with reasoning billed elsewhere. One bundles everything. One bills each direction separately, so your cost depends on how much each side talks. Averaging them produces a figure no vendor publishes, and the headline that understates most is OpenAI's, precisely because the backend agent sits outside it.
The benchmark does not hand OpenAI a clean win
Worth being honest about this, because it cuts against the easy story.
On the Artificial Analysis Speech-to-Speech Index, Gemini 3.8 Live Extended Thinking scores 82.6, ahead of GPT-Live-1 at 81.5. Gemini takes the top line.
GPT-Live-1 wins somewhere narrower and, for a phone agent, more useful. On conversational dynamics, the sub-score closest to what a caller actually experiences, it reaches 94.9 percent, and the lower reasoning-effort configuration reaches 97.3 percent. It runs full duplex, shifts context part way through a call, and delivers acknowledgement cues while the caller is still speaking, which drops perceived latency below what a stopwatch shows.
OpenAI separately reports a 30 percentage point gain over GPT-Realtime-2.1 on Full Duplex Bench. That figure is theirs, published by them, and we could not verify it independently. Treat it as a vendor claim.
Voice cloning, and who can actually use it
Both OpenAI and xAI support custom voices. Getting to them is where they differ.
OpenAI gates custom voices to eligible customers through a sales conversation. Grok's console gives any user up to 30 cloned voices free, currently in the US only, though the programmatic endpoint is reserved for Enterprise plans. xAI also describes a consent step: the speaker reads a passphrase that is transcribed and matched in real time, and speaker embeddings from the passphrase and the full recording are compared. On their account that means you cannot clone from a pre-existing recording. Nobody outside xAI has tested whether the check holds.
What this means if you are building
None of these three can be self-hosted. They are all closed weight, so the most you get is same-region placement, not your own hardware. What you can own is the layer around the model, which is the orchestration, the call logic and the transcripts.
That layer is where we spend our time. Dograh is open source and self-hostable, it runs any of the three behind a phone number, and it charges no platform fee because you run it. The per-minute platform fee is the line item that usually dwarfs the model anyway, whether you're self-hosting or using a hosted platform.
Pick GPT-Live-1 if you are building a phone agent today. Then keep the key swappable, because the ranking here has already moved once and will move again.
Top comments (0)