Grok Voice 3.0: xAI's voice model tops new leaderboard

What's new

xAI has released Grok Voice 3.0. It is a so-called speech-to-speech model: You talk to it, it understands and thinks through what you want, can use other tools along the way and replies with speech. It is available in xAI's Voice Agent API, the access developers use to build voice assistants into apps and phone systems.

The model ranks No. 1 on Artificial Analysis' Speech-to-Speech Index with 86.9 points, up from 81.3 for its predecessor, Grok Voice 2.0. By comparison, OpenAI's GPT-Live-1 scores 83.2 and Google's Gemini 3.8 Live scores 82.6. The figures come from xAI and from an independent blog that reproduces the measurement from October 13, 2026.

What's smart

The model is good at keeping a natural conversation going. On Full Duplex Bench, which measures how a voice handles pauses, turn-taking and interruptions, it scores 97 percent. On speech-based reasoning (Big Bench Audio), it scores 98.5 percent. That means it can both keep up when you interrupt and solve tasks you give it by voice.

On agent tasks, where the voice has to carry out something for a customer, the model rose from 56.5 to 71.6 percent on the τ-Voice benchmark. xAI also reports 74 percent on ServiceNow's EVA-Bench, which the company calls the highest score for an enterprise voice agent. In Artificial Analysis' Speech Agent Arena, task success is 94.8 percent.

Response time, meaning the time to first audio, is 0.81 seconds. That is about 40 percent faster than GPT-Live-1 at 1.34 seconds, but slightly slower than Grok Voice 2.0, which came in at 0.70 seconds.

xAI says the model powers millions of daily conversations in Tesla cars and handles about 20,000 daily calls in Starlink's call centers, where roughly 70 percent are resolved without human help. Those are xAI's own figures.

Cheaper or better?

The price is $0.08 per minute of audio, equivalent to $4.80 per hour, plus $0.004 per text input. That is unchanged from Voice 2.0, so you get a better model at the same price. On Big Bench Audio, one task costs about $0.025, versus $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1. No monthly price for regular users has been given, since the model is sold through the API.

It is better overall, but not at everything. On the τ-Voice agent benchmark, GPT-Live-1 scores 74.5 percent, putting it ahead of Grok Voice 3.0 at 71.6. Gemini 3.8 Live comes in at 68.6. According to the independent blog, long-term stability, performance under heavy load and noise robustness have not been measured. It also points out that automatic upgrades can risk making something worse in an existing setup.

What it's good for

The model is aimed at businesses that want voice assistants for customer service, phone switchboards and similar tasks where the voice has to understand, act and speak naturally. Users of the alias grok-voice-latest are switched to the new model automatically, and Grok Voice 2.0 has been retired. According to OrcaRouter, the API is compatible with the OpenAI Realtime API protocol via WebSocket, so you can swap the base URL and key without rebuilding everything.

For Danish readers, there is an open question: According to xAI, the model supports multiple languages and handles different accents and background noise, but the sources do not say whether Danish is included, or when and how the service will be available in Denmark and the EU.