Grok Voice 3.0 Tops the Charts for xAI

xAI released Grok Voice 3.0 on Tuesday, a speech-to-speech model it describes as the most intelligent voice system it has built, and the independent scorekeepers appear to agree. On Artificial Analysis's Speech-to-Speech Index, the model landed in first place with a score of 86.9, ahead of OpenAI's GPT-Live-1 at 83.2 and Google's Gemini 3.8 Live at 82.6. The previous version, Grok Voice 2.0, had sat at 81.3.

What separates a speech-to-speech model from the voice assistants most people know is that it does away with the middle steps. There is no transcription into text, no separate language model, no conversion back into audio. The system listens and speaks directly, which is what lets it handle the untidy parts of real conversation: pauses, interruptions, someone talking over the top of an answer. xAI puts the response latency at under one second.

Built for the phones that never stop ringing

The benchmark that matters most commercially is not the headline index but the grind of customer support. Here the picture is more mixed. On a test xAI calls Tau-Voice, which throws accents, background noise and interruptions at the model, Grok Voice 3.0 scored 71.6 percent. OpenAI's GPT-Live-1 edged it at 74.5 percent, though xAI notes its own model ran with roughly 40 percent lower latency. On a ServiceNow enterprise benchmark it claimed the best combined score for accuracy and customer experience.

xAI is not speaking hypothetically about call centres. It says Grok Voice already handles around 20,000 customer calls a day for Starlink across several languages, resolving about 70 percent without a human and converting one in four sales calls into a purchase. A single agent, by the company's account, verifies the caller, troubleshoots the hardware, issues credits and arranges replacement kit. The model also runs the in-car voice features in Tesla vehicles, where xAI says it powers millions of conversations daily.

Priced to spread

The pricing is the quietly aggressive part. Grok Voice 3.0 costs eight cents a minute, or 4.80 dollars an hour, exactly the same as the model it replaces. Developers who were pointing at the grok-voice-latest alias were moved across automatically, with no code changes required, though xAI suggests production users pin the explicit version name to avoid surprises when the next upgrade lands.

The launch fits a wider pattern this autumn, with the big labs racing to turn capable models into agents that do real work rather than answer questions. Google has been pushing an agent that labours for days at a time, and Cohere has built an enterprise platform that runs agents across different models. Voice is the frontier where the quality gap is most audible, and for the moment xAI holds the top line on the chart.