Grok Voice 3.0: benchmarks and what changes on the phone
Grok Voice 3.0 is xAI's new voice model, released on 14 October 2026. According to the official announcement, it ranks first on Artificial Analysis' speech-to-speech index with 86.9 points. For anyone using an AI phone assistant that is good news, for three reasons: more natural conversations, faster replies and a lower cost per task than its rivals. It does not win every test, though: on simulated support calls it still trails OpenAI's GPT-Live-1.
This article covers what was actually released, what the benchmarks say and what it means for a business that lets an AI phone assistant answer its calls.
What xAI released, in short
- The model: Grok Voice 3.0, called
grok-voice-3.0in the API. It is a speech-to-speech model: it listens and replies in voice directly, without transcribing to text and synthesising again. Fewer steps usually mean less waiting and a more natural tone. - Availability: live now in the Voice Agent API. Anyone using the
grok-voice-latestalias was moved to the new model automatically, and the API is unchanged (xAI release notes). - Usage price for developers: the same as Grok Voice 2.0.
- Languages and telephony: the documentation lists more than 20 languages with native-quality accents, and connections via SIP, WebSocket or LiveKit.
- Real-world use, as stated by xAI: millions of conversations a day in Tesla cars. Starlink's call centres use it for about 20,000 calls a day in several languages, with roughly 70% resolved without an agent.
A note on the company name: SiliconANGLE reports that xAI, acquired by SpaceX, is now called SpaceXSI (SiliconANGLE, 14 October 2026). The product is still Grok.
The benchmarks: where it wins and where it doesn't
Benchmarks are standardised tests: the same tasks are given to different models and the results are compared. This is the table published by xAI, with data from Artificial Analysis.
Grok Voice 3.0 compared: benchmarks reported by xAI
Benchmark What it measures Grok Voice 3.0 Grok Voice 2.0 GPT-Live-1 (Astra, Medium) Gemini 3.8 Live (Extended Thinking, High) Artificial Analysis Speech-to-Speech Index overall score 86.9 81.3 83.2 82.6 Big Bench Audio reasoning on spoken questions 98.5% 97.2% 90.1% 97.7% Speech Agent Arena (Elo) listener preference 1070 1011 1048 990 Task Success Rate tasks completed 94.8% 94.6% 87.4% 89.1% Full Duplex Bench pauses, interruptions, talking over 97.0% 95.1% 94.9% 91.9% τ-Voice support calls with accents and noise 71.6% 56.5% 74.5% 68.6% Best result in each row in bold. Source: xAI, Grok Voice 3.0 announcement, 14 October 2026, with Artificial Analysis data.
Three things worth noting:
- On the phone, the last two rows matter most. Full Duplex Bench measures how the model handles interruptions and pauses, which is what happens when a caller talks over it or changes their mind mid-sentence. Grok Voice 3.0 leads there with 97%. τ-Voice simulates support calls with accents and background noise, the test closest to a real phone call: the model jumps a long way from version 2.0 (56.5% to 71.6%), but GPT-Live-1 is still ahead.
- Speed. xAI claims latency under one second, and τ-Voice results comparable to GPT-Live-1 with 40% lower latency. It does not publish millisecond figures, though.
- Cost per task. According to Artificial Analysis, on Big Bench Audio Grok Voice 3.0 costs less per task than both GPT-Live-1 and GPT-Realtime-2.1. For phone assistant providers, that is margin to reinvest in longer calls or better reasoning models.
xAI also cites ServiceNow's EVA-Bench, more than 200 business scenarios, with 74% accuracy, the highest reported, but it does not publish competitor figures. These are all vendor-reported results: independent checks will follow in the coming weeks.
What changes for a call handled by an AI
For a small business, the model's name matters less than how the call goes. In practice, the possible changes are:
- Fewer awkward pauses. A model that answers in under a second and copes well with interruptions makes the call feel like a conversation rather than an answering machine.
- More calls resolved first time. The jump on τ-Voice suggests the model understands real requests better, accents and noise included. That means fewer "sorry, could you repeat that?".
- Multilingual by design, but worth testing in the field: regional accents, proper names and phone numbers are the tests that separate a good assistant from a mediocre one.
Who it matters most to
- Restaurants: calls come in during service, often with noise in the background. A model that handles noise and interruptions takes the booking without asking three times for the time and number of guests.
- Medical practices: older patients, unfinished sentences, requests that change mid-call ("actually, Thursday would be better"). Handling interruptions is decisive here.
- Auto repair shops: customers often call from the road or the car. Understanding speech in noisy places counts for more than the overall score.
What to do now
If you already use an AI phone assistant, you don't need to do anything: the model is a component your provider can upgrade behind the scenes. If you are evaluating a solution, the right questions are not about this month's model:
- Make a test call from a noisy place and interrupt the assistant mid-sentence.
- Ask which voice models the provider uses and whether it can switch when a better one comes out for your use case.
- Check what it does during the call: appointments, answers about your services, handover to a person, a summary afterwards. The features matter more than benchmarks.
On the yourang.ai home page you can listen to the voices of the providers available in the platform, Grok included, each one tested in a different kind of business. For a comparison of the options available in Italy, see our guide to the best AI phone assistants of 2026.
In short
Grok Voice 3.0 raises the bar for voice models, especially on handling interruptions and on speed, and confirms that the real-time voice race between xAI, OpenAI and Google benefits anyone using AI on the phone. On support calls, however, GPT-Live-1 is still ahead: a good phone assistant should be judged on your calls, not on a leaderboard. If you want to hear how an AI phone assistant answers in a case like yours, book a demo.
Frequently asked questions
What is Grok Voice 3.0?
It is xAI's new speech-to-speech model, released on 14 October 2026. It listens and answers in voice directly, without going through text, and it is built for voice agents: assistants that hold conversations on the phone or in an app.
Is Grok Voice 3.0 the best voice model?
It depends on the test. It ranks first on Artificial Analysis' speech-to-speech index (86.9, against 83.2 for GPT-Live-1 and 82.6 for Gemini 3.8 Live), but on the τ-Voice support-call benchmark it scores 71.6%, behind GPT-Live-1 at 74.5%.
Which languages does Grok Voice 3.0 speak?
xAI's documentation lists more than 20 languages with native-quality accents, Italian included. As with any model, real-world quality should be checked with test calls in your own language.
Can it be connected to a phone number?
Yes. According to xAI's documentation the API connects via SIP, WebSocket or LiveKit, and there is a sample phone agent built on Twilio. For a small business, though, the connection is handled by the phone assistant provider, together with the calendar, CRM and handover to a person.
