On October 14, SpaceXAI released Grok Voice 3.0, a voice model for agents that talk on the phone and in the car. It took first place in the Artificial Analysis Speech-to-Speech index: 86.9 points against 83.2 for OpenAI's GPT-Live-1. But "first place" depends on which table you open: with human voters the model is third, and in noisy calls OpenAI leads.
What this model is
Grok Voice 3.0 is a speech-to-speech model: it listens to speech and answers in voice directly, without converting to text and back. That means lower latency and more natural pauses and interruptions. It runs through the Voice Agent API, and the model ID is grok-voice-3.0.
The new version replaced Grok Voice 2.0 in the grok-voice-latest alias, so agents that pointed to it switched on their own. The vendor price is $0.08 per minute of conversation, the same as for 2.0. This is SpaceXAI's price, not GPTunneL's: our catalog has no such model.
What the measurements show
On the release page, SpaceXAI gives Artificial Analysis measurements from October 13. The index has four equal parts of 25%: speech understanding (Big Bench Audio), agent tasks (τ-Voice), human voting (Speech Agent Arena) and the share of completed tasks. Below are all the metrics from the vendor's table:
| Measure | Grok Voice 3.0 | Grok Voice 2.0 | GPT-Live-1 (Astra, Medium) | Gemini 3.8 Live (Extended Thinking, High) |
|---|---|---|---|---|
| Speech-to-Speech index | 86.9 | 81.3 | 83.2 | 82.6 |
| Big Bench Audio, speech understanding | 98.5% | 97.2% | 90.1% | 97.7% |
| Speech Agent Arena, rating from human voting | 1070 | 1011 | 1048 | 990 |
| Share of completed tasks | 94.8% | 94.6% | 87.4% | 89.1% |
| Full Duplex Bench, pauses and interruptions | 97.0% | 95.1% | 94.9% | 91.9% |
| τ-Voice, calls with noise and accents | 71.6% | 56.5% | 74.5% | 68.6% |
Bold marks the best result in a row. Of the four parts of the index, Grok wins three: speech understanding, human voting and the share of completed tasks. In the fourth, τ-Voice, it trails GPT-Live-1, and that part is the closest to real support calls.
Where "first place" does not hold
The vendor compares itself with specific variants of the models. In the table Gemini 3.8 Live is taken in extended thinking mode, and GPT-Live-1 is paired with GPT-6 Astra at medium effort. In the full Speech Agent Arena ranking the picture is different.
- Human voting. Grok Voice 3.0 with a rating of 1070 (±28) shares places 2 through 6. Above it by rating are Gemini 3.1 Flash Live Minimal (1100) and Gemini 3.8 Live in a different configuration than in the vendor's table (1088), but the intervals overlap, so you cannot say with confidence which is better.
- Noisy calls. In τ-Voice GPT-Live-1 leads: 74.5% against 71.6%. SpaceXAI replies that at comparable accuracy its model answers 40% faster, but that is the company's claim. SiliconANGLE's analysis points to this test as well.
- Share of completed tasks. Here Grok has the best result of all in the ranking, 94.8%.
What the company says
The other figures in the release are so far confirmed only by SpaceXAI itself:
- Starlink handles about 20,000 customer calls a day with Grok Voice in several languages, roughly 70% are resolved without a human, and one in four sales calls ends in a purchase.
- In Tesla the model holds millions of conversations a day: drivers check their calendar and email by voice.
- On EVA-Bench, ServiceNow's open benchmark for enterprise voice agents, it has the highest accuracy, 74%. The company gave no competitor figures.
Nobody has verified this data independently. That does not mean it is wrong, but it should not be taken as proof of quality either.
What this means for you
If you are building a voice agent for support, a ranking does not replace your own test. Take recordings of real calls with noise, accents and interruptions and check on them what the model understands and how it brings a task to the end. That is exactly the part the index shows worst.
If you just want to talk to AI by voice, the index will tell you little: it measures how agents work in support scenarios, not how pleasant an ordinary conversation is.
What we have for voice
There is no real-time voice agent like Grok Voice on GPTunneL yet. But you can voice a ready text with a neural network voice: the Text to Speech tool has dozens of voices, and you can clone your own from a recording of 15 seconds or more.
Sources
- Introducing Grok Voice 3.0, SpaceXAI, 14.10.2026 - release, price, Artificial Analysis measurements, company data.
- Grok Voice 3.0 in the documentation, SpaceXAI - model ID.
- Speech-to-Speech Index methodology, Artificial Analysis - what the index is made of.
- Speech Agent Arena, Artificial Analysis, as of 16.10.2026 - ranking by human voting.
- Grok Voice 3.0 outscores OpenAI and Google, SiliconANGLE, 14.10.2026 - independent analysis and caveats.